The Bleeding Edge

// Episode W18 · 2026-04-24 to 2026-05-01

GPT-5.5 launched, beat Opus 4.7 on benchmarks

GPT-5.5 launched, beat Opus 4.7 on benchmarks. OpenAI shipped GPT-5.5 seven days after Anthropic's Opus 4.7 — the two models traded leadership across coding, reasoning, and multimodal benchmarks in the same week. The Neuron's read: "the two models feel like they swapped souls." A…

Bleeding Edge -- Episode Briefing W18

Sources scanned: The Neuron Daily (x7), Superhuman/The Code, The New Stack Note: Only ralph.behnke91@gmail.com scanned. ralph@operatingmodel.ai not connected to Gmail MCP.

Headline of the Week

GPT-5.5 launched, beat Opus 4.7 on benchmarks. OpenAI shipped GPT-5.5 seven days after Anthropic's Opus 4.7 — the two models traded leadership across coding, reasoning, and multimodal benchmarks in the same week. The Neuron's read: "the two models feel like they swapped souls." A defining moment in the frontier-model arms race; the gap between Anthropic and OpenAI shrunk to days, not months.

Consolidated Stories by Category

Frontier & Big Tech (12 stories)

  1. GPT-5.5 launched, beat Opus 4.7 on benchmarks -- Seven days after Opus 4.7, OpenAI shipped GPT-5.5. The Neuron says "the two models feel like they swapped souls." [The Neuron, 2026-04-24] [MULTI-NEWSLETTER]

  2. Anthropic raising $40-50B at $900B valuation -- If confirmed, this would be one of the largest private fundraises in history. [The Neuron, 2026-05-01]

  3. OpenAI cut Microsoft loose / Apple-style stack is the new endgame -- OpenAI restructured its relationship with Microsoft, pursuing a vertically integrated Apple-style approach. [The Neuron, 2026-04-28]

  4. Anthropic now lives inside all four hyperscalers -- Claude available on AWS, Azure, GCP, and Oracle. Earnings day coverage. [The Neuron, 2026-04-30]

  5. Elon Musk testimony: xAI distilled OpenAI models, $97.4B bid for OpenAI assets -- Musk admitted xAI partly distilled OpenAI models, called his $38M early donation "I was a fool," disclosed a $97.4B Musk-led bid. Greg Brockman testifies next. [The Neuron/CNBC, 2026-04-30] [MULTI-NEWSLETTER]

  6. OpenAI restricted GPT-5.5-Cyber after slamming Anthropic for the same -- OpenAI rolled out GPT-5.5-Cyber to vetted "critical cyber defenders" while Anthropic launched Claude Security in public beta. Both putting frontier capabilities in defenders' hands. [The Neuron, 2026-05-01]

  7. DeepSeek V4 shipped with 98% KV cache reduction -- Compressed Sparse Attention + Heavily Compressed Attention slashes memory on long-context tasks. [The Neuron, 2026-05-01]

  8. Big labs "pulling the ladder" on distillation -- Clement Delangue (Hugging Face) argues labs that used distillation to build their empires now use lawyers to stop competitors doing the same. Same week Musk admitted xAI distilled OpenAI. [The Neuron, 2026-05-01] [MULTI-NEWSLETTER]

  9. Demis Hassabis: West needs strong open-source AI to beat China -- Google's CEO argues US risks losing without it, edge models should be open-source since they're already exposed on-device. [The Neuron, 2026-05-01]

  10. One analyst spent $6K/day on Claude, replaced 100-person team -- Meta also axed 10% of workforce to feed AI investment. [The Neuron, 2026-04-26]

  11. Codex for Work shipped to enterprise -- OpenAI rolled out workplace capabilities. Aaron Levie at Box started hiring "agent engineers." [The Neuron/Superhuman, 2026-05-01] [MULTI-NEWSLETTER]

  12. Anthropic analyzed 1M Claude conversations -- Found 6% are personal guidance. Sycophancy hit 25% in relationship conversations. Opus 4.7 cut that rate in half. [The Neuron, 2026-05-01]

Apps / Dev Tools / Platforms (7 stories)

  1. Cursor's $60B bet: the harness is the product, not the model -- Shipped Cursor SDK, harness team published long read, SpaceX/xAI Colossus partnership for training proprietary models. Google told The New Stack it doesn't care which coding tool devs use. [The New Stack, 2026-05-01] [MULTI-NEWSLETTER]

  2. OpenClaw powered by a 4-tool coding agent called Pi -- WhatsApp-based personal AI assistant runs on an open-source tool with only read, write, edit, and bash. Mario Zechner's "Slow the F*** Down" got standing ovations. Armin Ronacher (Flask creator) found code quality dropping industry-wide: "vibe slop." [The Neuron, 2026-05-01]

  3. ElevenLabs launched ElevenMusic -- The $11B voice generation lab enters AI music. Joins Suno and Udio. [Superhuman, 2026-05-01]

  4. Suno crossed $300M ARR with 2M paid subscribers -- AI music going from fad to full industry. Breaking Rust became first AI artist on Billboard #1. Xania Monet landed $3M deal from Suno. [Superhuman, 2026-05-01]

  5. Poolside shipped M.1 and Laguna XS.2 coding models -- Free on OpenRouter, handling 10B+ tokens/day. [The New Stack, 2026-05-01]

  6. OpenRouter Owl Alpha -- Stealth high-performance model optimized for agentic workloads, 1M context, free to try. [The Neuron, 2026-05-01]

  7. Best Value AI 2026 -- Compares 37+ LLMs by quality-adjusted tokens per dollar. Updated April 2026. [The Neuron, 2026-05-01]

Infrastructure & Ecosystem (3 stories)

  1. AWS Bedrock shaping Model Context Protocol -- Luca Chang discussed Amazon's open-source MCP contributions at MCP Summit NYC. [The New Stack, 2026-05-01]

  2. Cursor agents get kanban board -- Cursor now lets you control agents like a project manager, tracking progress on a board. [Superhuman, 2026-05-01]

  3. Claude Code push notifications -- Now sends push to your phone when long tasks complete or need input. Pair mobile Claude app. [The Neuron, 2026-05-01]

People in AI (3 stories)

  1. Theo Browne called Anthropic an "evil cult" -- Prominent developer/YouTuber publicly urged engineers to resign. 1.2M views. [Superhuman, 2026-05-01]

  2. Matt Pocock's AI Engineer workshop -- 256K views on real workflow for AI coding: grill-me alignment, vertical-slice tracer bullets, AFK loops, deep-module architecture. [The Neuron, 2026-05-01]

  3. Victor Taelin: re-explaining domain knowledge is the AI dev nightmare -- AGENTS.md, RAG, SKILLs, fine-tuning all fail for unknown unknowns. Nightly fine-tuning on your domain is the missing product. [The Neuron, 2026-05-01]

AI Gone Wrong / Harms (2 stories)

  1. "Vibe slop" is the new technical debt -- Armin Ronacher (Flask) found after interviewing 30+ engineering teams that code quality has dropped across the industry. Agents generate garbage future agents can't process. [The Neuron, 2026-05-01]

  2. Amazon product description podcasts -- Amazon turning product descriptions into podcasts. Major backlash. 1.3M views. [Superhuman, 2026-05-01]

Skills / Prompting (1 story)

  1. Prompting rules changed for both GPT-5.5 and Claude 4.7 -- Claude 4.7 went literal (does exactly what you type). GPT-5.5 went autonomous (drop the step-by-step scripts). Both penalize vague prompting but from opposite directions. [The Neuron, 2026-05-01] [MULTI-NEWSLETTER]

AI & Robotics (2 stories)

  1. Generalist Gen-1 robot ties zip ties with improvisation -- Robot lost grip mid-task, used other hand to readjust. "Improvisational intelligence in action." [The Neuron, 2026-05-01]

  2. AGIBOT Finch: 16-robot fleet learning while deployed -- Making cocktails, restocking groceries, brewing Gongfu tea. Learning from real-world tasks. [The Neuron, 2026-05-01]

AI in Consumer Hardware (1 story)

  1. Mira smart glasses act as a "second brain" -- Monitor your day, build profile from conversations, remember preferences, book appointments, translate 60+ languages. [Superhuman, 2026-05-01]

Investment (1 story)

  1. DeepSeek price war -- DeepSeek aggressive pricing putting pressure on pricing across the industry. AI costs now exceeding salary costs for some teams. [The Neuron, 2026-04-27]

Regions / Macro (1 story)

  1. The disappearing AI middle class -- OpenAI and DeepSeek pricing bets have split the market. Developers must adapt to a new economy. [The New Stack, 2026-05-01]

Research (1 story)

  1. Human Creativity Benchmark -- Claude best for ideation, Gemini leads design systems, ChatGPT best at refinement. No model leads all three phases. [Superhuman/Contra Labs, 2026-05-01]

Enterprise Adoption (1 story)

  1. Executives vibe-coding their own tools -- CEOs using AI coding tools to build agents, dashboards, production systems without developers. "I was tired of explaining it to somebody who was supposed to build it for me." [The New Stack, 2026-05-01]

Story Counts by Category

Category Count
Frontier & Big Tech 12
Apps / Dev Tools / Platforms 7
Infrastructure & Ecosystem 3
People in AI 3
AI Gone Wrong / Harms 2
AI & Robotics 2
Skills / Prompting 1
AI in Consumer Hardware 1
Investment 1
Regions / Macro 1
Research 1
Enterprise Adoption 1
Total 35

Multi-Newsletter Stories (high signal)

  • GPT-5.5 vs Opus 4.7 (Neuron + Superhuman)
  • Elon Musk testimony / xAI distillation (Neuron + Superhuman)
  • Distillation ladder-pulling (Neuron -- connects Musk testimony + Delangue commentary)
  • Codex for Work (Neuron + Superhuman)
  • Cursor harness bet (New Stack + Superhuman)
  • Prompting rules changed (Neuron + Superhuman)

Notes

  • This was an unusually dense week -- GPT-5.5 launch, Musk trial, Anthropic mega-raise, Cursor's strategic pivot all in 7 days
  • Only 1 of 2 Gmail accounts scanned -- ralph@operatingmodel.ai newsletters (Innovating with AI, Human+Agent Daily, EU Digital & Tech, Google Cloud AI, Replit) not included
  • The "vibe slop" / Pi / OpenClaw cluster is a strong deep-dive candidate
  • The Anthropic $900B raise + distillation hypocrisy + Theo Browne "evil cult" is another potential thread

User Additions

Must-include: AI Harness explainer (supports Cursor story #13)

Explain what an "AI harness" is for the audience. Cursor's bet is that the model becomes a commodity and the harness — the orchestration layer around it (tool calling, context management, agent loops, file handling, MCP integration) — is the real product. This connects to the multi-LLM orchestration deep dive: if the harness is model-agnostic, you can swap models without rebuilding your workflow.

Deep Dive: Memory + Multi-LLM Orchestration (2-segment episode)

Full deep dive research provided by host. Content below.

L2 RESEARCH OUTPUT

Segment 1: Humor

FACT ANCHORS

  1. Anthropic is raising $40-50B at a $900B valuation — up from $380B in February.
  2. Cursor struck a $60B deal with SpaceX, betting the harness around AI models is the real product.
  3. Elon Musk admitted under oath that xAI "partly distilled" OpenAI's models to build Grok.
  4. OpenAI launched GPT-5.5, codename "Spud," one week after Anthropic's Opus 4.7.
  5. Claude 4.7 went literal (does exactly what you type) while GPT-5.5 went autonomous (figures it out itself) — both penalize vague prompts, from opposite directions.

THE SET

  1. [contrast] Anthropic is now valued at nine hundred billion dollars. Their annual revenue is thirty billion. For context, that's a 30x multiple. Coca-Cola trades at six. Investors are basically saying Claude is five Coca-Colas of optimism.

  2. [translation] What Anthropic said: "We're committed to building safe, beneficial AI." What Anthropic's cap table says: "We're committed to building AI so fast that we need forty billion dollars in fresh capital three months after the last round."

  3. [contrast] Elon Musk testified under oath that xAI "partly distilled" OpenAI's models. In the same trial where he's suing OpenAI for betraying its mission. The man copied their homework and then reported them to the principal.

  4. [character beat] OpenAI launched GPT-5.5, codename Spud. One week after Opus 4.7. Somewhere at Anthropic, a product manager is staring at a launch calendar and whispering "can we not have one week."

  5. [escalation] Claude 4.7 now does exactly what you type. No more guessing what you meant. GPT-5.5 does the opposite — you describe the goal, it figures out the path. So one model stopped reading between the lines, and the other started writing between them. Your prompts from six months ago now fail on both, for completely different reasons.

  6. [contrast] Cursor is valued at sixty billion dollars. Their thesis: AI models are becoming commodities and the harness is the product. SpaceX agreed so hard they wrote a ten-billion-dollar check. Google's response was essentially "we don't care which coding tool developers use." Which is either strategic patience or the world's most expensive shrug.

  7. [specific callback] One analyst spent six thousand dollars a day on Claude and replaced a hundred-person economics team. Meta's response: cut 10% of staff to fund more AI. The machines haven't taken your job. They've taken your department's budget.

SHAREABLE LINES

  • "Investors are saying Claude is five Coca-Colas of optimism."
  • "He copied their homework and then reported them to the principal."
  • "One model stopped reading between the lines, the other started writing between them."

WRITER NOTES

  1. Anthropic/Coca-Cola — Story #2. Mechanic: scale absurdity (30x vs 6x multiple). Safe.
  2. Anthropic translation — Story #2. Mechanic: corporate translation. Safe.
  3. Musk distillation — Story #5. Mechanic: contrast (suing + copying). Safe — factual record from testimony.
  4. Spud timing — Stories #1, #2. Mechanic: character beat. Safe.
  5. Prompting split — Story #28. Mechanic: escalation. Safe.
  6. Cursor/SpaceX — Story #13. Mechanic: contrast + tag. Safe — "expensive shrug" is the risk line, defensible as commentary.
  7. 100-person team — Story #10. Mechanic: specific callback. Edgy-but-defensible — factual, names no individual.

Segment 2: Top 5

1. Anthropic raising $40-50B at a $900B valuation

Anthropic is in talks to raise $40-50 billion in fresh capital at a valuation between $850 billion and $900 billion, more than doubling its $380 billion valuation from February. Annual revenue run rate has surged past $30 billion (up from ~$9 billion at end of 2025). Google plans to invest up to $40 billion, and Amazon has committed up to $25 billion with access to 5 gigawatts of compute. Why it matters: This would be one of the largest private fundraises in history, signaling that AI companies are now valued like nation-state infrastructure, not software startups. Corroborated

2. Cursor's $60B bet: the harness is the product, not the model

Cursor shipped its SDK, published its harness architecture, and struck a deal with SpaceX to train proprietary models on xAI's Colossus supercomputer. CEO Michael Truell declared this the "third era" of AI development. Google told The New Stack it doesn't care which coding tool developers use. 70% of Fortune 1000 companies now use Cursor, with $6B+ annualized revenue projected by year-end. Why it matters: The entire industry — Anthropic, OpenAI, Google, Microsoft — now agrees the harness is the product. They just disagree on what to charge. Corroborated

3. Elon Musk admits xAI distilled OpenAI models under oath

During federal testimony in his lawsuit against OpenAI, Musk confirmed xAI "partly" used OpenAI's models to train Grok via distillation. He called his original $38M donation "I was a fool" and disclosed a $97.4B Musk-led bid for OpenAI's assets. Same week, Hugging Face CEO Clement Delangue accused big labs of "pulling the ladder" on distillation — using it to build their empires, then lawyering up to stop competitors. Why it matters: First high-profile public confirmation of model distillation between domestic rivals. Changes the distillation debate from theoretical to evidenced. Corroborated

4. GPT-5.5 launched — codename "Spud"

OpenAI released GPT-5.5 on April 23, seven days after Anthropic's Opus 4.7. The model excels at code, research, data analysis, and multi-tool agentic workflows. Benchmarks show improvements over Opus 4.7 on Terminal-Bench 2.0 (82.7%) and FrontierMath. API access withheld until April 24 for "different safeguards." Available to Plus, Pro, Business, and Enterprise — not free tier. Why it matters: The model race is now measured in days, not months. And GPT-5.5's strengths (autonomy, tool use) vs Opus 4.7's strengths (precision, literalness) mean the models are diverging in philosophy, not just capability. Corroborated

5. Prompting rules changed for both Claude 4.7 and GPT-5.5 — from opposite directions

Claude 4.7 went literal: does exactly what you type, no longer compensates for fuzzy intent. GPT-5.5 went autonomous: drop the step-by-step scripts, describe the outcome, let the model pick the path. Both penalize vague prompting, but the fix is opposite — be surgically specific for Claude, be goal-oriented for GPT. Anthropic and OpenAI both published new prompting guides this month. Why it matters: Every listener's existing prompts now underperform on at least one of the two dominant models. This is the most practically actionable story of the week. Corroborated

Segment 3: Categorised News

Frontier & Big Tech

Anthropic now available on all four hyperscalers — Claude is now accessible via Amazon Web Services (AWS), Microsoft Azure, Google Cloud Platform (GCP), and Oracle Cloud. This came alongside Q1 earnings season where all four hyperscalers reported accelerating AI infrastructure spend. Why it matters: Anthropic's multi-cloud strategy eliminates vendor lock-in as a reason not to adopt Claude in enterprise. Corroborated The Neuron (2026-04-30)

OpenAI restructured its Microsoft relationship — OpenAI is pursuing a vertically integrated Apple-style approach, reducing dependence on Microsoft while building its own hardware, search, and distribution capabilities. Why it matters: The OpenAI-Microsoft partnership that defined the first AI wave is unwinding. OpenAI wants to own the stack. Corroborated The Neuron (2026-04-28)

OpenAI restricted GPT-5.5-Cyber; Anthropic launched Claude Security — Both companies put frontier-model cybersecurity capabilities in defenders' hands. OpenAI rolled out to vetted "critical cyber defenders" while Anthropic went public beta. Why it matters: AI-powered offensive security is already here. These are the first dedicated defensive products. Corroborated The Neuron (2026-05-01)

Anthropic analyzed 1M Claude conversations — Found 6% are people seeking personal guidance. Sycophancy hit 25% in relationship conversations. Opus 4.7 cut that rate in half vs 4.6. Why it matters: Anthropic is publicly measuring and fixing sycophancy — a problem most AI companies ignore or deny. Corroborated Anthropic Research (2026-04-30)

DeepSeek V4 shipped with 90-98% KV cache reduction — Compressed Sparse Attention (CSA) plus Heavily Compressed Attention (HCA) slash memory requirements for long-context tasks. V4-Pro requires only 27% of the inference compute and 10% of the KV cache of V3.2. Why it matters: DeepSeek is making 1M-token contexts practical at a fraction of the cost — compressing the pricing floor for every competitor. Corroborated MarkTechPost (2026-04-24)

Big labs "pulling the ladder" on distillation — Hugging Face CEO Clement Delangue argued that labs which used distillation to build their empires now deploy lawyers to stop competitors doing the same — the same week Musk admitted xAI did exactly that. Why it matters: Distillation is the AI industry's open secret. It just got confirmed under oath. Corroborated The Neuron (2026-05-01)

Demis Hassabis: the West needs open-source AI to beat China — Google DeepMind's CEO argues the US risks losing the AI race without a strong open-source stack, and that edge models should be open-source because once deployed on-device they're already exposed. Why it matters: Google's AI chief making the open-source case is significant — and contradicts Google's own history of keeping models proprietary. Corroborated The Neuron (2026-05-01)

Apps / Dev Tools / Platforms

OpenClaw powered by a 4-tool coding agent called Pi — The WhatsApp-based AI assistant (250,000+ GitHub stars) runs on Mario Zechner's open-source Pi framework: just read, write, edit, and bash. Armin Ronacher (Flask creator) found code quality dropping industry-wide, coining "vibe slop." Central thesis: agents don't feel pain, so they generate code future agents can't process. Why it matters: The anti-complexity movement in AI tooling has a working proof of concept — and a pointed critique of agent-swarm approaches. Corroborated The Neuron (2026-05-01), Pragmatic Engineer

ElevenLabs launched ElevenMusic — The $11B voice generation lab entered AI music with a platform combining generation, discovery, and remixing. Built on a "fully licensed music model." Currently free, 7 songs/day. Why it matters: ElevenLabs entering means AI music has a serious well-funded player with a licensing-first approach, unlike Suno/Udio which face active RIAA lawsuits. Corroborated TechCrunch (2026-04-02)

Suno crossed $300M ARR (Annual Recurring Revenue) with 2M paid subscribers — Breaking Rust became the first fully AI artist to hit #1 on a Billboard chart. Telisha Jones used Suno to create Xania Monet and landed a $3M record deal. Why it matters: AI music has crossed the novelty threshold into a real industry with real revenue and real chart positions. Rights-holder economics becomes the next battleground. Corroborated TechCrunch (2026-02-27)

Codex for Work shipped to enterprise — OpenAI rolled out document, spreadsheet, and slide capabilities plus workplace app integrations. Box CEO Aaron Levie immediately started hiring "agent engineers." Why it matters: "Agent engineer" as a job title signals the shift from using AI tools to orchestrating AI workflows. Corroborated The Neuron / Superhuman (2026-05-01)

Best Value AI 2026 — Compares 37+ Large Language Models (LLMs) by quality-adjusted tokens per dollar across local hardware, APIs, and subscriptions. Updated April 2026. Why it matters: The first serious attempt at per-dollar LLM comparison — practically useful for anyone deciding between providers. Corroborated Desktop Commander (2026-04)

AI & Robotics

Generalist Gen-1 robot improvises mid-task — Robot lost grip on a zip tie, used its other hand to readjust, then completed the task. Research lead called it "improvisational intelligence in action." Why it matters: Robots recovering from unexpected failures without explicit programming is a step-change from scripted manipulation. Corroborated The Neuron (2026-05-01)

AGIBOT Finch: 16-robot fleet learning while deployed — Fleet autonomously improves from real-world tasks including making cocktails, restocking groceries, and brewing Gongfu tea. Why it matters: Learning-while-deployed is the manufacturing equivalent of continuous deployment in software. Corroborated The Neuron (2026-05-01)

AI in Consumer Hardware

Mira smart glasses act as a "second brain" — Monitor conversations, build a profile of preferences, book appointments, send emails, translate 60+ languages. Why it matters: The first consumer hardware that attempts persistent personal memory through ambient listening. Privacy implications are significant and largely unaddressed. Corroborated Superhuman (2026-05-01)

People in AI

Theo Browne called Anthropic an "evil cult" — Prominent developer and YouTuber publicly urged Anthropic engineers to resign. 1.2M views. Why it matters: Developer sentiment toward AI companies is fracturing. The safety narrative that once shielded Anthropic from criticism is no longer working with this audience. Unverified — single source (X post), no secondary reporting on specific claims.

One analyst replaced a 100-person economics team using Claude at $6K/day — Simultaneously, Meta cut 10% of its workforce to fund AI investment. Why it matters: The "AI replaces teams" narrative moved from hypothetical to documented case study. The $6K/day number makes the ROI calculation concrete. Unverified — single newsletter report, no primary source documentation.

Infrastructure & Ecosystem

AWS Bedrock shaping Model Context Protocol (MCP) — At the MCP Summit in New York City, AWS Bedrock's Luca Chang discussed Amazon's open-source contributions to MCP. Why it matters: MCP is becoming the standard for how AI agents connect to tools. AWS's involvement signals enterprise readiness. Corroborated The New Stack (2026-05-01)

Segment 4: Prompting Skill

Dual-Model Prompting: How to Write for Both Claude 4.7 and GPT-5.5

Best for: Anyone who uses both Claude and ChatGPT and noticed their prompts stopped working as well in April 2026.

Steps:

  1. Define the outcome first — Before you open either tool, write one sentence: "Success looks like [X]." Both models now reward this.

  2. For Claude 4.7 — be surgically specific. List every variable. If you want 3 paragraphs, say 3 paragraphs. If you want bullet points, say bullet points. Claude no longer infers format, scope, or intent from vague instructions.

  3. For GPT-5.5 — describe the goal, not the process. Drop the step-by-step scripts. Say "analyze this dataset and tell me what's interesting" rather than "Step 1: load the CSV. Step 2: compute mean..."

  4. Use this structure for both (adapted from OpenAI's published guidance):

Role: [what the model is]
Goal: [the outcome you need]
Success criteria: [what must be true]
Constraints: [limits]
Output: [format, length, tone]
  1. Test your prompts on the wrong model. If your Claude prompt works on GPT-5.5, it's probably too vague for Claude. If your GPT prompt works on Claude, it's probably over-specified for GPT.

Example prompt (Claude 4.7):

You are a financial analyst. Summarize this earnings report in exactly 5 bullet points. Each bullet: one metric, one trend direction, one sentence of context. No headers. No preamble. Output only the 5 bullets.

Example prompt (GPT-5.5):

You are a financial analyst. I've attached our Q1 earnings report. Tell me the 3 things our board should be worried about and the 2 things they should celebrate. Be blunt.

Common failure: Using the same prompt style for both models. Claude 4.7 with a vague prompt gives narrow, literal, sometimes useless output. GPT-5.5 with an over-specified prompt gives mechanical, checkbox-y output.

Fix: Maintain two prompt templates. Or use the universal structure above and adjust specificity per model.

Variants:

  • Effort-level variant (Claude only): Set reasoning effort to medium for routine tasks, xhigh for complex analysis. This is the new temperature.
  • Messy-input variant (GPT-5.5 only): Paste raw meeting notes, data dumps, or multi-part questions. GPT-5.5 is designed to handle mess. Claude 4.7 needs it cleaned up first.
  • Chain variant: Use Claude 4.7 for structured first-pass extraction, then feed the output to GPT-5.5 for synthesis and insight.

Segment 5: New AI Tools

1. Best Value AI 2026

What it does: Compares 37+ LLMs by quality-adjusted tokens per dollar across local hardware, APIs, and subscriptions. Who for: Anyone deciding between AI providers or wondering if they're overpaying. Why now: Updated April 2026 with empirical quota tests. First tool to normalize across subscriptions, APIs, and local models in one view. Quick workflow: Visit the site, filter by your use case (coding, writing, analysis), sort by value score, compare against what you're currently paying. Source: desktopcommander.app/best-value-ai

What it does: Chat with legal documents for verbatim citations, draft contracts, run spreadsheet-style tabular reviews across hundreds of files with every cell linked to a page and quote. Who for: Legal teams, solo practitioners, anyone reviewing contracts or compliance documents. Why now: Self-hostable with your own Claude or Gemini keys. No data leaves your infrastructure. Quick workflow: Deploy on your server, connect your API key, upload documents, ask questions in natural language. Every answer links to exact page and quote. Source: mikeoss.com

3. OpenRouter Owl Alpha

What it does: Stealth high-performance foundation model optimized for agentic workloads with 1M context window and strong tool use. Who for: Developers building agent systems who want a strong, cheap alternative to frontier models. Why now: Free to try during alpha. Processing 10B+ tokens/day already. Quick workflow: Use via OpenRouter API. Set as your default model in Claude Code Router or any OpenRouter-compatible tool. Source: openrouter.ai/openrouter/owl-alpha

Segment 6: AI Personality

Mario Zechner — The Developer Who Said "Slow the F*** Down" and Built OpenClaw's Engine

Who: Austrian developer, creator of the libGDX game framework, and builder of Pi — the minimal 4-tool coding agent that powers OpenClaw (250,000+ GitHub stars).

What this week: Sat down with The Pragmatic Engineer for 90 minutes alongside Armin Ronacher (Flask creator). Central argument: agent armies create complexity their own future selves can't untangle. His blog post "Slow the F*** Down" got standing ovations at AI Engineer Europe.

Why they matter: While the industry races to build agent swarms, Zechner built the most-starred repo on GitHub with four tools: read, write, edit, bash. His bet — that the personalization layer of every AI tool will converge toward Pi's minimalism within two years — is the most provocative contrarian take in AI tooling right now.

Safe fun fact: Pi users can ask Pi to modify Pi itself. Non-engineers have done this with zero coding skills. The tool rewrites its own source code as a feature, not a bug.

Sources: The Pragmatic Engineer, mariozechner.at

Segment 7: Catch-all

AI Music Crossed the Novelty Threshold — and Nobody's Ready for the Rights Fight

Suno has $300M ARR and 2M paid subscribers. Breaking Rust became the first fully AI artist to hit #1 on Billboard. A woman in Mississippi used Suno to turn her poetry into an AI-generated song and landed a $3M record deal. ElevenLabs just entered the market with ElevenMusic, built on a "fully licensed" music model — pointedly differentiating from Suno and Udio, which face active RIAA lawsuits alleging "mass infringement."

Why this is the catch-all: AI music is no longer a novelty or a demo. It's an industry with chart-toppers, record deals, and licensing battles. The next 12 months will determine whether AI music creators get rights-holder economics or get sued into oblivion. ElevenLabs' licensing-first approach vs Suno's scale-first approach is the test case.

Corroborated TechCrunch (2026-02-27), Music Ally (2026-04-30)

Deliverable 1: Show Notes (bullets only)

  • Anthropic in talks to raise $40-50B at $900B valuation — revenue run rate past $30B, up from $9B end of 2025
  • Cursor ships SDK, strikes $60B SpaceX deal — thesis: models are commodities, the harness is the product
  • Musk admits under oath xAI distilled OpenAI models to build Grok — same trial where he's suing them
  • GPT-5.5 launches codename "Spud," one week after Opus 4.7 — model race now measured in days
  • Prompting rules changed: Claude 4.7 went literal, GPT-5.5 went autonomous — your old prompts fail on both
  • DeepSeek V4 cuts KV cache by 90-98% with new attention architecture
  • OpenClaw's secret: a 4-tool coding agent called Pi, 250K+ GitHub stars
  • Armin Ronacher coins "vibe slop" — code quality dropping industry-wide from agent overuse
  • ElevenLabs enters AI music with ElevenMusic (licensing-first); Suno at $300M ARR
  • Breaking Rust: first AI artist to hit #1 on Billboard
  • Mira smart glasses: ambient AI memory for daily life
  • Demis Hassabis: West needs open-source AI to beat China
  • One analyst replaced 100-person economics team at $6K/day Claude spend
  • Agent engineers: Box CEO hiring for a role that didn't exist 6 months ago
  • Best Value AI 2026: first per-dollar LLM comparison across 37+ models
  • DEEP DIVE: AI agent memory (5-layer model) + multi-LLM orchestration ($60/dev vs $200)

Deliverable 2: Blog Summary (~1,000 words)

The Week AI Became Infrastructure

The week of April 24-May 1, 2026 will be remembered as the week AI stopped being a product category and started being valued like infrastructure. Three numbers tell the story: $900 billion, $60 billion, and $6,000 per day.

The Nine-Hundred Billion Dollar Question. Anthropic is raising $40-50 billion at a valuation that would top $900 billion — more than doubling from $380 billion just three months ago. Annual revenue has surged past $30 billion, up from $9 billion at end of 2025. Google is investing up to $40 billion; Amazon has committed $25 billion plus 5 gigawatts of compute. These aren't software company numbers. These are nation-state infrastructure numbers. Inference If Anthropic closes at $900B, the combined private valuation of the top three AI labs (OpenAI, Anthropic, xAI) exceeds the GDP of most European countries. That level of capital concentration in three companies building the same technology has no historical precedent.

The Harness, Not the Model. Cursor's $60 billion SpaceX deal crystallized a thesis the industry has been circling: the AI model is becoming a commodity, and the product is the harness — the orchestration layer that handles tool calling, context management, agent loops, and file handling. 70% of Fortune 1000 companies now use Cursor. Google told The New Stack it doesn't care which coding tool developers use. The models compete on capability; the harness competes on workflow. For listeners wondering what "harness" means in practice: it's everything between you and the model. When you use Claude Code or Cursor or Copilot, the model generates text, but the harness decides what files to read, when to call tools, how to handle errors, and when to ask for your input. Cursor's bet is that getting this layer right matters more than which model sits underneath.

The Distillation Admission. Elon Musk admitted under oath that xAI "partly distilled" OpenAI's models to train Grok — during the very trial where he's suing OpenAI for betraying its nonprofit mission. The same week, Hugging Face's CEO accused big labs of "pulling the ladder" on distillation, using the technique to build their empires and then deploying lawyers to stop competitors. Distillation — where a smaller model learns from a larger model's outputs — is the AI industry's open secret. This week it became a matter of legal record.

The Prompting Split. Both Claude 4.7 and GPT-5.5 shipped new behavior this month, and both punish the same habit (vague prompting) from opposite directions. Claude 4.7 went literal: it does exactly what you type, no longer inferring intent from fuzzy instructions. GPT-5.5 went autonomous: describe the goal, let the model figure out the path. If you're using both models, you now need two prompting styles. This is the most practically actionable development of the week.

The $6,000/Day Question. One analyst reportedly spent $6,000 per day on Claude and replaced a 100-person economics team. The same week, Meta cut 10% of its workforce to fund AI investment. Inference The economics of AI replacement are becoming concrete enough to model: $6K/day for Claude vs the fully loaded cost of 100 economists. Even with skepticism about the specific numbers, the directional math is compelling enough to change headcount planning in every large organization.

The Music Industry's Reckoning. Suno hit $300M ARR with 2M paid subscribers. Breaking Rust became the first AI artist to top Billboard. ElevenLabs entered the market with a licensing-first model. AI music is no longer a demo — it's an industry with chart positions and record deals. The next 12 months will determine whether creators get rights or get sued.

What Ties It Together. The connecting thread across all of these stories is commoditization pressure. Models are commoditizing (Cursor's thesis). Prompting is commoditizing (both labs published free guides). Music creation is commoditizing (anyone with Suno can chart). What isn't commoditizing: taste, judgment, orchestration, and the discipline to use these tools well. Inference The value is shifting from "having access to AI" to "knowing when to use which AI, and when to stop."

Deliverable 3: Short Skill Article (~500 words)

Your Old Prompts Are Broken: How to Write for Claude 4.7 and GPT-5.5

If you use both Claude and ChatGPT, you may have noticed your prompts getting worse results since mid-April 2026. You're not imagining it. Both Anthropic and OpenAI changed how their models handle instructions — and they went in opposite directions.

Claude 4.7 went literal. It now does exactly what you type. If you write "summarize this," you get a summary — and nothing else. No suggestions, no additional context, no inferred formatting. The model that used to read between the lines now reads only the lines.

GPT-5.5 went autonomous. It handles messy, multi-part tasks by planning its own approach. The step-by-step scripts that worked on GPT-5.4 now produce mechanical, checkbox output. Describe the goal; let the model choose the path.

The universal fix: Start every prompt with what success looks like.

For Claude: be specific about every variable — format, length, tone, what to include, what to exclude. Shorter prompts, but more precise.

For GPT-5.5: describe the outcome and constraints, then get out of the way. Longer context is fine; prescriptive process steps are counterproductive.

The structure that works for both (adapted from OpenAI's published guide):

Role: [what the model is]
Goal: [the outcome]
Success criteria: [what must be true]
Constraints: [limits]
Output: [format, length, tone]

Claude reads every field literally. GPT reads the goal and success criteria, then improvises the rest. Same structure, different interpretation — and that's fine.

One test to calibrate: Take your best Claude prompt and run it on GPT-5.5. If it works perfectly, it's probably too vague for Claude. Take your best GPT prompt and run it on Claude. If it works perfectly, it's probably over-specified for GPT.

Deliverable 4: Meme

Caption: "The AI model race, April 2026"

Image-gen prompt: A cartoon-style illustration of two runners on a track. The runner labeled "Claude" is reading a very detailed instruction manual while running. The runner labeled "GPT" has thrown away the manual and is sprinting freestyle. Both are tripping over the same hurdle labeled "vague prompts." A crowd of tiny confused users watches from the stands. Bright colors, clean lines, exaggerated expressions. No real-person likenesses.

Alt caption 1: "One follows the instructions too literally. The other ignores them entirely. Your prompts fail on both."

Alt caption 2: "Claude: 'You said summarize, so I summarized.' GPT: 'You said summarize, so I restructured your entire business.'"

Deliverable 5: Weekly Inferences

  1. Inference AI company valuations have decoupled from revenue multiples entirely. Anthropic at 30x revenue and Cursor at 10x revenue are priced on future monopoly position, not current earnings. This is either visionary or the next bubble.

  2. Inference The model-as-commodity thesis is now consensus among the major players, which means the next competitive battleground is the agent harness layer — tool calling, context management, multi-step orchestration.

  3. Inference Musk's distillation admission will accelerate the legal and regulatory framework around model training data rights. Expect ToS enforcement to tighten within 90 days.

  4. Inference The prompting split between Claude (literal) and GPT (autonomous) reflects a deeper philosophical divergence: Anthropic is building a precision instrument, OpenAI is building an autonomous agent. These will serve different markets within 12 months.

  5. Inference "Agent engineer" as a job title signals that AI tool use is becoming a specialized skill, not a general capability. Companies that treat AI as "everyone can use it" will underperform those that hire dedicated orchestrators.

  6. Inference The Pi/OpenClaw phenomenon (4 tools, 250K stars) is the strongest signal yet that developer sentiment is turning against complexity. The next wave of successful AI tools will be radically simple.

  7. Inference AI music's $300M ARR and Billboard #1 positions make the rights fight inevitable and imminent. ElevenLabs' licensing-first approach is a bet that the music industry's legal machinery will crush unlicensed competitors within 18 months.

  8. Inference Mira's "second brain" glasses represent the inevitable collision between persistent AI memory and privacy law. The first regulatory action against ambient AI recording will come from the EU within 12 months.

  9. Inference The 100-person economics team replacement story — even if exaggerated — changes the narrative from "AI augments workers" to "AI replaces departments." HR and finance leadership will start modeling AI replacement costs against headcount in Q3 2026.

  10. Inference DeepSeek V4's 98% KV cache reduction makes million-token contexts economically viable. This will collapse the "long context is too expensive" argument within 6 months, opening new application categories that were previously cost-prohibitive.

Self-Check

  • Date range respected (Apr 24 - May 1, 2026)
  • Sources present on every story
  • Verification labels correct (Corroborated with 2+ sources, Unverified where single-source)
  • No duplicates across segments
  • Top 5 are genuinely the top 5 by audience impact
  • Short but complete
  • Humor + shareable lines included
  • First use of abbreviations spelled out (ARR, AWS, GCP, MCP, LLM, KV, CSA, HCA, RIAA)

DEEP DIVE: MEMORY + MULTI-LLM ORCHESTRATION

Format: Two-segment episode Segment 1: Memory — why agents forget, and what serious teams are doing about it Segment 2: Multi-LLM orchestration — how to pair Claude with cheaper models without breaking everything Connecting thread: Memory is the substrate that makes multi-LLM orchestration possible. Without shared memory, every model swap is a context reset.

How to Use This Document

This is the consolidated research base for the episode. It is structured so that Claude Code (or another agent) can:

  1. Read the full document for context
  2. Help develop specific segments — show notes, talking points, scripted intros, transcripts, follow-up content
  3. Stress-test the arguments — find weak claims, missing counterpoints, places where the evidence is thinner than the assertion
  4. Generate derivative content — Twitter threads, LinkedIn posts, newsletter writeups, YouTube descriptions

When working on the episode, treat this as a living document. Add interview notes, listener questions, and fresh research at the bottom under "Working Notes."

The voice of the episode should be candid and specific. Not "AI is changing everything" — the listener already knows that. The value is in the failure modes, the trade-offs, the things that look right but break under load.

SEGMENT 1: MEMORY

The Cold Open Hook

Every coding agent in 2026 still has the "50 First Dates" problem. You can have a four-hour productive session with Claude Code, watch it learn your codebase, build genuine momentum — and tomorrow morning it starts from zero. The instruction file you wrote helps. The auto-memory it accumulates helps. But the fundamental amnesia is still there, and it gets worse the moment a second developer or a second tool enters the picture.

The clever framing: this isn't really a memory problem. It's five different memory problems pretending to be one. And conflating them is why most teams' "AI memory strategy" is a single bloated CLAUDE.md file that everyone has stopped reading.

The Core Claim

Memory for coding agents is five distinct concerns. Each has its own home, its own format, and its own update cadence. Skipping any one creates a category of pain the others can't fix.

Layer What it holds Where it lives Cadence
L1: Project rules Stack, conventions, "always do X" AGENTS.md in repo root, committed Weekly / on-PR
L2: Tool-specific config MCP setup, slash commands, tool features CLAUDE.md, GEMINI.md, .github/copilot-instructions.md Rare
L3: Personal context Individual preferences, sandbox URLs *.local.md (gitignored), user-global ~/.<tool>/... When it bugs you
L4: Task & dependency state What to work on, what's blocked, who claimed what Beads (.beads/) committed to repo Every session
L5: Cross-session knowledge Decisions, learned facts, "we tried X and it broke" Mem0 (or Zep/Letta) via MCP Continuous

L1 + L2 are instructions. L3 is personal overrides. L4 is work state. L5 is episodic knowledge. You need all five.

The Architectural Decisions That Fall Out

1. AGENTS.md is the source of truth. It's now governed by the Linux Foundation's Agentic AI Foundation. Codex CLI, Copilot, Cursor, Windsurf, Aider, Devin, Warp, and Antigravity (v1.20.3+) read it natively. Claude Code and Gemini CLI need a one-line bridge file or a symlink. One file. No duplication.

2. Symlink the tool-specific files to AGENTS.md. CLAUDE.md, GEMINI.md, WARP.md, .github/copilot-instructions.md — all symlinked. The moment you maintain two files with overlapping content, drift starts and the whole stack rots. This is the single highest-leverage architectural decision.

3. Beads for task state, not Claude Code Tasks. Claude Code Tasks (introduced 2026, explicitly inspired by Beads) is excellent but locked to Claude Code. With multiple agents in play, you need git-synced state visible to all of them.

4. Mem0 for cross-session knowledge. Has an MCP server, works with every tool, most mature option. Zep is the alternative for temporal reasoning. Letta if you want to rebuild around memory entirely.

5. One repo, one standard file tree. Every repo gets the same scaffolding. Make a create-project script. Treat any deviation as a bug.

The Two Categories Most People Conflate

Task state != knowledge memory. This is the most important distinction in the entire space and it gets blurred constantly.

  • Task state (Beads): "bd-42 is blocked by bd-37, claimed by Alice's agent, discovered from bd-19"
  • Knowledge memory (Mem0): "We tried Redis Streams for the queue and it broke under load"
  • Instructions (AGENTS.md): "Use pnpm. Tests required for new code. Don't edit packages/generated/."

Most "AI memory" products solve the second category and call it "memory." That's true but partial. The first category is what actually changes whether agents can pick up where they left off across sessions and across each other.

Beads -- The Honest Take

Beads is Steve Yegge's CLI issue tracker that gives AI coding agents persistent task state across sessions. It stores work as a dependency-aware graph in a Dolt SQL database synced via JSONL files in git. Hash-based IDs prevent merge collisions. bd ready --json returns unblocked work in priority order.

What it offers that no built-in agent memory does:

  • Queryable dependency graph
  • Cross-tool persistence (any agent reads it)
  • Atomic claim semantics for multi-agent work
  • Team-shared work state via git
  • Audit trail through git history
  • Discovery linkage (--deps discovered-from:bd-X)

The limitations nobody warns you about:

Agents don't reach for it unprompted. Yegge himself calls it "a leaky abstraction." Instructions in AGENTS.md lose weight as context grows mid-session. Mitigation: install the beads-mcp MCP server so agents see it as a first-class tool, not a CLI to forget. Wire bd sync into git hooks. Build "land the plane" as a single command.

Sync conflicts and occasional data loss. Yegge has publicly documented sessions where issues vanished during merges. Unpushed work in multi-agent setups causes severe conflicts. Mitigation: bd doctor --fix on a schedule. Treat git push as part of done. Use Dolt server mode for real concurrent writes.

Scale ceiling around 500 open issues. When agents read issues.jsonl directly with jq, they hit ~25K token limits. Mitigation: bd admin compact regularly. Force agents to use bd ready --json not cat issues.jsonl.

Project maturity. Yegge himself describes the architecture as "crummy by pre-AI standards, requires AI to work around its edge cases." Daily releases. Single-maintainer bus factor. Mitigation: pin versions, upgrade weekly not daily.

Wrong tool for wrong work. Beads is for "current week" — not PRDs, not roadmaps. Stuffing roadmap items destroys bd ready signal.

Granularity is on you. Issues over ~2 minutes of work cause agent context rot mid-task.

The meta-point that's worth saying out loud on the show: every Beads limitation has a workaround, but most workarounds are human discipline, not tooling. If your team will enforce "land the plane" rituals and run bd doctor regularly, Beads earns its keep. If you're hoping to install it and forget about it, you'll get burned.

The Pilot Plan

Setup: One repo. 2-3 developers who already use Claude Code or equivalent heavily. Decide upfront: replacement for your tracker, or agent-only memory layer that mirrors to Jira? Two weeks.

Week 1: Single-agent workflows. Each dev runs sessions using bd ready to pick work and "land the plane" to close. Granularity audit on day 5.

Week 2: Stress tests deliberately designed to trigger known failure modes:

  • Concurrent claim test (two devs claim same issue within a minute)
  • Cross-session handoff (Dev A files issues, Dev B picks them up next day)
  • "Forgot to push" recovery (deliberately end without git push, see how painful resolution is)
  • Data loss audit (compare bd list --status all --json | jq 'length' across machines — must match)

The single most predictive metric: unprompted query rate. If agents aren't reaching for bd ready --json on their own by end of week 2, no amount of tooling will save the rollout.

Go/no-go: Green at >=70% unprompted queries, zero unrecovered data loss, >=90% land-the-plane completion. Red on any unrecovered data loss.

Alternatives Worth Naming

For task state: Claude Code Tasks (great but locked to Claude Code, ~70% of Beads' value), GitHub Issues + gh CLI (works for humans but lacks ready semantics), Linear/Jira via MCP (pragmatic if your org already runs them).

For knowledge memory: Mem0 (most mature, MCP server), Zep with Graphiti (temporal knowledge graph, ~15 points better on LongMemEval for time-sensitive reasoning), Letta formerly MemGPT (rebuild around memory entirely; Letta Code shipped March 2026), Anthropic's Memory Tool (managed /memories directory but only via API, not Claude Code).

Segment 1 Talking Points

  • The five-layer model — and why one bloated CLAUDE.md is the single most common failure mode
  • The symlink trick: maintain one AGENTS.md, point everything else at it
  • Why Beads matters specifically for teams (auto-memory is per-machine; Beads is git-synced)
  • Yegge's own admission that Beads is "a leaky abstraction" — the discipline tax is real
  • The pilot plan listeners can actually run
  • The unprompted query rate as the single best signal
  • Why Mem0 != Beads != CLAUDE.md (this confusion is endemic)

Segment 1 Questions Worth Asking on the Show

  • What's the strongest argument against this stacked approach? (Complexity tax. Solo dev on one project is fine with just CLAUDE.md.)
  • When does Claude Code Tasks make Beads obsolete? (Probably within 12 months for single-tool teams.)
  • Is there a world where vector DBs or RAG replace this? (No — vector retrieval is for unstructured corpus; this is structured state.)
  • What's the team-size threshold where this becomes worth the discipline tax? (~3 devs concurrently with agents, or solo dev across ~5+ active projects.)

Segment 1 Pitfalls to Call Out

  • Bloating AGENTS.md (over 500 lines, Codex CLI silently truncates)
  • Auto-generated AGENTS.md from /init commands — generic and noisy
  • Duplication across CLAUDE.md, GEMINI.md when symlinks would prevent it
  • Treating Mem0 like a wiki (it's for atomic facts, not documentation)
  • Letting one dev skip the standard scaffolding — divergence kills team consistency
  • Linting rules in AGENTS.md (that's Prettier's job)
  • Forgetting git push — Beads explicitly defines work as "not done until pushed"

SEGMENT 2: MULTI-LLM ORCHESTRATION

The Cold Open Hook

Anthropic's Claude Code Max plan is $200/month. Z.AI's GLM Coding Plan is $18. The GLM model self-reports 94.6% of Claude Opus's coding performance. Independent benchmarks suggest 75-85% on real-world tasks. So the question listeners are quietly asking: do I really need to keep paying Anthropic?

The honest answer is more interesting than yes or no. The right framing is: which 20% of your work needs Opus, and which 80% can run on something a tenth the cost? Get that allocation right and you cut your spend by 60-70% while preserving quality where it actually matters. Get it wrong and you spend three hours undoing a confident-but-wrong architectural answer from a cheap model that should never have been asked the question.

The Connecting Thread to Segment 1

Multi-LLM orchestration is impossible without the memory stack from segment 1. When you swap from Claude Opus to GLM-5.1 mid-task, the new model gets nothing the old one learned during the session except what's in visible context. KV caches don't transfer. Auto-memory doesn't transfer. But Beads state does transfer because it's a queryable graph in git. Mem0 does transfer because it's a remote MCP server any model can hit. AGENTS.md does transfer because it's project state, not session state.

The memory stack isn't just a nice-to-have alongside multi-model routing. It's what makes multi-model routing functional rather than chaotic. Without it, every model swap is a context reset.

The Three Patterns That Actually Work

Pattern 1: Subscription substitution. Replace Anthropic Max with Z.AI's GLM Coding Plan. Set ANTHROPIC_BASE_URL and ANTHROPIC_MODEL env vars and Claude Code talks to GLM. Skills, MCP, subagents, hooks all keep working. Best when you want one cheaper model for everything.

Pattern 2: Routed hybrid via Claude Code Router (CCR). @musistudio/claude-code-router sits between Claude Code and any combination of providers. Standard four-tier routing:

  • default -> GLM 4.7 (everyday work)
  • background -> DeepSeek (silent tool calls and file scans)
  • think -> Kimi K2 Thinking (multi-step reasoning)
  • longContext -> DeepSeek or local Gemma (>100K tokens)

Pattern 3: Tier-by-phase. Anthropic Pro for the hard 20%. GLM Coding Plan for the daily 80%. Don't try to route automatically — devs switch tools consciously. Total ~$40/month vs. $200/month Max.

What Each Model Is Actually Good At (April 2026)

Model Best for Avoid for Cost
Claude Opus 4.6/4.7 Hard architecture, gnarly debugging, long agentic chains Routine implementation $$$$
Claude Sonnet 4.6 General default if paying Anthropic anyway — $$$
GLM-5.1 Coding-heavy daily work, refactors, near-Opus quality Frontier reasoning, novel domains $
GLM-4.7 "Competent junior dev" workhorse Architecture decisions $
Kimi K2 / K2.5 / K2 Thinking Multi-step reasoning, long-horizon planning Speed-critical interactive work $$
DeepSeek V3.2 / V4 Cheap bulk work, file scans, background tasks Anything requiring nuance cents
Qwen 3.5/3.6 Long context (1M+), Chinese-language work Frontier coding $

The Caveats That Belong on Air

  • Vendor benchmarks lie politely. GLM-5.1's "94.6% of Opus" is self-reported using Claude Code as the testing harness — home-field advantage.
  • Claude Code is not actually model-agnostic. Same model performs better through OpenCode or Cline than through Claude Code.
  • The 16K system prompt cache trap. Non-caching providers multiply input costs by 4-5x.

Failure Modes That Bite

  • Don't break tool calling (MCP servers stop working with malformed JSON)
  • Don't switch models mid-conversation when context is dense
  • Don't violate Z.AI's ToS (GLM Coding Plan restricted to supported tools)
  • Don't cheap out on architecture decisions — hard rule: design questions go to Opus
  • Don't forget to set ANTHROPIC_DEFAULT_HAIKU_MODEL
  • Don't bet a deadline on it before a month of low-stakes use

The Recommended Stack -- and the Real Numbers

Per developer: ~$60/month total

  • Anthropic Pro (~$20) for the hard 20%
  • GLM Coding Plan Lite ($18) for the daily 80%
  • OpenRouter prepaid ($20 buffer) for overflow

vs. $200 for Max alone. Quality on the bottom 80% essentially indistinguishable.

Segment 2 Talking Points

  • The "20/80 allocation" framing
  • The three orchestration patterns
  • Why GLM's "94.6% of Opus" is technically true and practically misleading
  • Claude Code is not actually model-agnostic
  • The economics: ~$60/dev vs. $200/dev
  • How memory makes orchestration functional rather than chaotic

Segment 2 Questions Worth Asking on the Show

  • Where's the inflection point at which paying Anthropic stops making sense entirely?
  • Will Anthropic respond by cutting prices, or leaning into capabilities GLM/Kimi can't match?
  • What happens when DeepSeek V4 ships at $0.30/million tokens with claimed 80%+ SWE-bench?
  • Is there a security/compliance angle for Western enterprises sticking with Anthropic?

Segment 2 Pitfalls to Call Out

  • Trusting vendor self-reported benchmarks
  • Assuming Claude Code is fully model-agnostic
  • Cheaping out on architecture decisions
  • Forgetting ANTHROPIC_DEFAULT_HAIKU_MODEL
  • Switching models mid-task instead of at boundaries
  • Running CCR on critical client work before a month of low-stakes use

EPISODE-LEVEL CONNECTIVE TISSUE

The single sentence: Memory is what turns a collection of cheap models into a coherent team; without it, you're paying less to lose more.

Recurring metaphor: The 50 First Dates problem.

Contrarian take: Most teams don't have a memory problem — they have a discipline problem. Adding more memory tools to a team that doesn't use the ones they have just adds entropy.

Optimistic take: A solo developer with a serious memory stack and intelligent routing operates at the productivity of a 3-person team from 2024. The compounding is real, the discipline tax is real, and most teams haven't figured this out yet — which is the opportunity.

SHOW PREP CHECKLIST

  • Decide on episode length (45-min and 90-min versions both work)
  • Identify guest (Yegge for segment 1; CCR production user for segment 2)
  • Pull live demo: bd ready --json, CCR routing config, Mem0 dashboard
  • Verify benchmarks and prices are current as of recording date
  • Sketch cold open and tease for segment 2 at end of segment 1
  • Decide whether to publish companion blog post with configs and commands

DERIVATIVE CONTENT IDEAS

  • Twitter thread (segment 1): "The five layers of AI agent memory" — 8-12 tweets
  • Twitter thread (segment 2): "How to cut your AI coding bill by 70%" — 20/80 framing
  • LinkedIn post: Frame around team-level economics
  • Newsletter writeup: Long-form with file paths and configs
  • YouTube short: The 16K system prompt cache trap, 60 seconds
  • GitHub repo: The onboard-dev.sh and AGENTS.md template

WORKING NOTES

Add interview transcripts, listener questions, fresh research below this line.

DEEP DIVE: AI IN THE XIAOMI DRAGON CHASSIS

Format: Single-segment deep dive Topic: How Xiaomi used AI to build the most intelligent production car chassis — and what it signals about AI moving from screens to steel Added: 2026-05-02 (L3 deep dive)

The Cold Open Hook

A phone company just shipped the most AI-dense car chassis in production. Not Tesla. Not Mercedes. Not BMW. Xiaomi — the company most people know for $300 smartphones — put 700 TOPS of AI compute, a unified robot-and-car brain, and predictive road-scanning suspension into a sedan that starts at $31,870. It sold 15,000 units in 34 minutes.

The reframing: the Dragon Chassis isn't really a suspension upgrade with AI bolted on. It's a robotics platform that happens to have wheels. And the reason that matters is because Xiaomi built the same AI model that controls their humanoid robots and deployed it into a car. The chassis doesn't just react to the road — it reasons about what the road will do next, using the same spatial intelligence that lets a robot tie zip ties.

The Core Claim

Xiaomi's Dragon Chassis is the first production vehicle where the autonomous driving AI and the physical chassis control share a single foundation model. Every competitor treats ADAS (Advanced Driver-Assistance Systems) and chassis dynamics as separate systems with separate brains. Xiaomi unified them through MiMo-Embodied — an open-source model that bridges robotics and driving — and a new architecture called XLA that replaces rule-based lane-keeping with genuine spatial reasoning.

The AI Stack (What's Actually Running)

Layer Component What It Does
Compute NVIDIA Thor-U (700 TOPS) Runs both ADAS and chassis control on one chip — 8x more than the previous Orin chip (84 TOPS)
Foundation Model MiMo-Embodied Cross-embodied vision-language model trained on both robotics and driving data. SOTA on 29 benchmarks
Cognitive Layer XLA Architecture Replaces end-to-end driving with multimodal reasoning — vision, audio, radar, nav data fused in latent space
Chassis Intelligence Dragon Chassis Controller AI-based road preview, slip detection, predictive suspension adjustment
Sensors 11 cameras + LiDAR + 4D mmWave radar + 12 ultrasonics Feed both ADAS and chassis systems simultaneously
Domain Architecture Four-in-One Domain Control Consolidates driving, chassis, cabin, and connectivity onto unified compute

What Makes XLA Different from "End-to-End" Driving

Traditional end-to-end autonomous driving (Tesla's approach): train a neural network to go from camera pixels to steering commands. Fast, but opaque. When it fails, nobody knows why.

Xiaomi's XLA approach: the "X" stands for cross-modal. It fuses vision, audio, radar, and navigation data — then reasons in latent machine language (not text, not pixels). The key distinction, per Xiaomi's VP of autonomous driving Chen Long: "XLA can combine on-site signs with environmental information, understand that this is a road closure detour scenario, and intelligently reroute. End-to-end systems would continue forward."

Three core capabilities:

  1. Spatial perception — Centimeter-level precision (robotics data gives it this; driving-only models are decimeter-level)
  2. Status prediction — Forecasts what other road agents will do next
  3. Driving planning — Generates safe maneuvers with explainable justifications

The embodied intelligence angle: because MiMo-Embodied is also trained on robotic manipulation data, the car AI understands physical consequences. Chen Long: "Spatial reasoning is really about helping the car understand what consequences a certain driving choice could produce."

How AI Controls the Chassis (Not Just the Driving)

This is the part most coverage misses. The AI doesn't just steer — it physically prepares the suspension for what's coming.

Road Preview System

The Dragon Chassis uses the car's cameras and LiDAR to scan road surface conditions ahead of the vehicle. Combined with cloud-sourced road data (think: crowdsourced pothole maps), the system predicts surface changes before the car reaches them and pre-adjusts the suspension. Xiaomi calls this "intelligent chassis preview with lift functionality."

How it works in practice:

  • Camera + LiDAR detect a pothole 15+ meters ahead
  • AI classifies severity and predicts optimal suspension response
  • Dual-chamber air springs with CDC (Continuous Damping Control) pre-soften before impact
  • Result: what Xiaomi calls "zero-bump driving" — the occupants never feel what the wheels hit

Context: Mercedes pioneered this concept with Magic Body Control in the S-Class (2013) — stereo cameras scan 15m ahead, hydraulic suspension compensates. But Mercedes' system is purely reactive pattern-matching. Xiaomi's is running the same spatial reasoning model that plans driving maneuvers. The AI doesn't just see a bump; it understands the bump in context of speed, tire grip, passenger comfort targets, and upcoming road geometry.

AI Slip Detection and Traction Control

The Dragon Chassis includes:

  • Coordinated traction control — AI coordinates all four motors (in quad-motor variants) 500 times per second per wheel
  • Dedicated wet/slippery-road mode — Not just reduced power; actively monitors grip via multimodal sensors
  • AI multimodal monitoring for slippery surfaces — Uses camera + radar + tire feedback to detect grip loss before the driver notices
  • Predictive chassis adjustment — If the AI sees wet road ahead, it pre-tensions the suspension and adjusts torque distribution before the car reaches the wet patch

The Fully Active Suspension (Pre-Research Tech)

Xiaomi also showed what's coming next (not yet in the 2026 SU7, but announced as pre-research):

  • 4.6 kW power per wheel
  • 140mm height adjustment range
  • Adjustment speed 100x faster than traditional air springs
  • "Zero bump, zero roll, zero pitch" target
  • Camera + cloud road preview for advance adjustment

This is where it connects to robotics: the suspension actuators are essentially robot limbs with enough power and speed to actively cancel road input, not just dampen it.

AI-Designed Materials: The "Material Genome" That Invented a New Alloy

This is the second AI story hiding inside the Dragon Chassis — and arguably the more radical one. Before the AI drives the car, it designed the metal the car is made from.

The Problem

Traditional aluminum die-casting requires heat treatment — a slow, energy-intensive process where cast parts are baked at high temperatures for hours to achieve structural strength. This is a bottleneck for production speed and a major cost driver. Tesla's Gigacasting solved the geometry problem (fewer parts), but still needed heat treatment to make the aluminum strong enough.

What Xiaomi Did

In collaboration with China's National Key Materials Laboratory, Xiaomi built an AI simulation system they call the "Material Genome" method. It evaluated over 10.16 million alloy formulas computationally to find one that achieves high structural strength without heat treatment.

Think of it as a generative AI for metallurgy. Instead of generating text or images, it generates alloy compositions — simulating mechanical properties, thermal behavior, castability, and cost for each combination — then selects the optimal candidate from millions.

The Result: Xiaomi Titan Alloy

Composition (from patent filings): Aluminum base with 0.3-3.5% Manganese, 0.4-2.0% Iron, 0.02-0.6% Silicon, 0.01-0.6% Chromium, 0.03-0.45% Titanium, 0.01-2.8% Nickel, 0.01-0.4% Vanadium, 0.01-0.5% Zirconium, up to 2.5% Zinc, 0.01-7.0% Rare Earth elements, plus microelements.

Properties achieved:

  • 17% lighter than conventional aluminum castings
  • No heat treatment required (eliminates hours of baking per part)
  • Higher crash resistance than traditional die-cast aluminum
  • 840 fewer welding points in the final structure
  • 2dB better cabin noise reduction (structural dampening)

Xiaomi claims to be "the only domestic car manufacturer with mass-produced, self-developed alloy materials."

The Manufacturing: Hypercasting

The Titan Alloy feeds into Xiaomi's 9,100-ton Hypercasting machine — exceeding Tesla's most advanced 9,000-ton Gigacasting press by 100 tons. The machine:

  • Weighs 1,050 tons and occupies 840 square meters
  • Merges 72 stamped-and-welded components into a single cast structure
  • Casts individual chassis sections in ~100 seconds
  • Uses a 5-zone, 8-gate mold design for complex geometries
  • Produces one complete car every 76 seconds

The system includes AI-driven parameter optimization during casting, sealed aluminum liquid automation with precision delivery, and computer vision quality inspection of every cast part.

Why This Matters (The AI Angle)

Traditional materials science: A PhD student tests 50-100 alloy compositions over 2-3 years. Each requires physical samples, lab testing, iterative refinement.

Xiaomi's approach: An AI system evaluated 10 million+ formulas computationally, found the optimal composition, and delivered a production-ready alloy that eliminates an entire manufacturing step. This is the same pattern as AlphaFold (protein folding), drug discovery AI, and battery materials research — but deployed at mass-production scale in a consumer product.

The compounding effect: AI designed the alloy → the alloy eliminates heat treatment → elimination speeds production → faster production enables the $32K price → the price enables 15,000 orders in 34 minutes → the volume funds more AI R&D. This is what vertical integration looks like when AI sits at the materials science layer, not just the software layer.

Limitations and Trade-offs

  • Repairability nightmare. A single integrated cast structure means minor collision damage may require replacing the entire rear section. Insurance premiums and repair costs rise.
  • No independent verification of the "10 million formulas" claim. Computational materials science at this scale is plausible (it's how battery makers work), but Xiaomi hasn't published the methodology.
  • Lock-in. Titan Alloy is proprietary. If supply chain disruptions hit the rare-earth elements, there's no drop-in alternative.
  • Recyclability questions. Complex multi-element alloys are harder to recycle than standard aluminum grades. At scale, this creates an end-of-life problem.

Sources

The MiMo-Embodied Model: Why It Matters

MiMo-Embodied is Xiaomi's open-source foundation model (released on Hugging Face and GitHub) that does something no other production model does: it handles both autonomous driving AND robotic manipulation in one architecture.

Architecture

  • Type: Cross-embodied vision-language model
  • Training: Progressive four-stage pipeline — embodied + driving skill learning → chain-of-thought inference → fine-grained reinforcement learning
  • Benchmarks: SOTA on 17 embodied AI benchmarks (task planning, affordance prediction, spatial understanding) + 12 autonomous driving benchmarks (perception, prediction, planning)
  • Key finding: "Capabilities learned in one domain enhance performance in the other" — robotics data makes the car AI better at spatial reasoning; driving data makes the robot AI better at navigation

Why Cross-Embodied Training Matters for a Car

A robot that ties zip ties understands force, grip, spatial relationships, and consequence at a resolution that pure driving data can't provide. When that same model runs in a car, it doesn't just see "obstacle ahead" — it reasons about the physical interaction between tire and surface, between suspension force and body roll, between steering input and vehicle trajectory.

Chen Guang (Xiaomi autonomous driving exec): "There are very few companies that can deploy such a complex model on an actual vehicle and then push it out to all users."

The Honest Take

What Works

  • Price-to-intelligence ratio is unprecedented. 700 TOPS, LiDAR, predictive suspension, unified AI architecture — starting at $31,870. A Mercedes S-Class with Magic Body Control is $120,000+. The BMW 7 Series with Executive Drive Pro is $100,000+. The tech gap is closing; the price gap isn't.
  • Unified compute saves weight and cost. One Thor-U chip replaces what used to be 3-4 separate domain controllers. Fewer chips = fewer wiring harnesses = lighter = more range.
  • Open-source model is a strategic power move. MiMo-Embodied being fully open allows the research community to validate and improve it. It also pressures competitors to open their driving models or fall behind on transparency.
  • 15,000 orders in 34 minutes. The market voted.
  • Embodied intelligence transfer is genuinely novel. No other production car benefits from robotics training data in its driving model.

What Doesn't (or Hasn't Been Proven Yet)

  • The Electrek first-drive couldn't test any of it. The test car was "still calibrating its system" — only self-parking was functional. We have Xiaomi's claims and benchmark numbers, but zero independent validation of XLA in real driving conditions.
  • Camera-based road preview has known limits. Night, rain, snow, construction zones, fresh potholes not in the cloud database. Mercedes' Magic Body Control had the same problem — works brilliantly on well-mapped roads, degrades in edge cases.
  • China-first deployment means China-trained AI. Road behavior, signage, driving norms are dramatically different in Europe and the US. When Xiaomi launches in Europe (planned 2027), the model needs retraining on entirely different road cultures.
  • Latent-space reasoning is opaque. XLA reasons in "machine language" for latency reasons, but this makes it harder to audit, debug, or explain failures. Explainability is claimed but unverified.
  • Regulatory barriers. NVIDIA Thor-U and LiDAR-based city NOA are legally permitted in China. European and US regulations are years behind. The car may arrive in export markets with capabilities software-locked.
  • Lei Jun's $8.7B AI bet is partially funded by car sales. If the SU7 margin compresses under price competition (Tesla just cut Model 3 again), the AI R&D budget faces pressure.

The Competitive Landscape

Maker Suspension AI Driving AI Shared Model? Price
Xiaomi SU7 2026 Road preview + predictive CDC + slip detection XLA/MiMo-Embodied (700 TOPS) Yes — unified $31,870
Tesla Model 3 Passive (no air suspension) FSD (HW4, ~300 TOPS) No $38,990
Mercedes S-Class Magic Body Control (camera road scan) Drive Pilot L3 (limited) No $120,000+
BMW 7 Series Executive Drive Pro (camera preview) Highway Assist No $100,000+
BYD Seal DiSus-C adaptive DiPilot (dual Orin) No $28,000
NIO ET7 Active suspension NIO Pilot (4x Orin, 1016 TOPS) No $55,000

The gap: Xiaomi is the only one with a single model controlling both systems. Everyone else runs driving AI and chassis dynamics as separate software stacks talking over CAN bus.

Actionable Takeaway

For the Audience (Non-Technical, AI-Curious)

What this means for you: The car you buy in 2027-2028 will have a robot brain, not just a computer. The chassis won't just absorb bumps — it'll predict them. The driving system won't just see lanes — it'll understand physics. And the phone company will offer this at half the price of the German incumbents.

What to watch for:

  • Independent reviews of XLA in real-world driving (expect mid-2026 from Chinese automotive press)
  • Xiaomi's Europe launch timeline and which features survive regulatory localization
  • Whether Tesla responds with a unified model or keeps chassis control separate
  • The open-source ecosystem around MiMo-Embodied — if universities start publishing improvements, the compounding is real

For the Show

Go/no-go on the story: Go. This is a "robotics meets automotive" story the audience hasn't heard framed this way. The phone-company angle and the price gap make it viscerally interesting even to non-car people.

Alternatives Worth Naming

  • Tesla's approach: Pure vision, no LiDAR, passive suspension on Model 3/Y. Bet on scale and data volume over sensor density. Cheaper hardware, bigger fleet.
  • Mercedes' approach: Camera-scanned suspension since 2013, but driving AI and chassis AI remain completely separate. More real-world data on road scanning, less integration.
  • Huawei's ADS 3.0: Similar Chinese tech stack (LiDAR + vision), partners with multiple OEMs. Not vertically integrated like Xiaomi.
  • NIO: Most compute (1016 TOPS) but hasn't unified driving and chassis models. Focus on battery-swap ecosystem instead.

Talking Points for the Show

  1. "A phone company shipped a robot brain in a car for $32K. That's the headline."
  2. The transfer learning angle: robot dexterity data makes the car better at understanding physics. Nobody else is doing this.
  3. The Mercedes comparison: Magic Body Control was $120K luxury tech in 2013. Xiaomi democratized it and added AI reasoning in 13 years, at a quarter of the price.
  4. The 700 TOPS compute point: the previous SU7 had 84 TOPS on its base model. This is an 8x jump in one model year. Moore's Law doesn't explain this — architectural unification does.
  5. Lei Jun's $8.7B AI investment isn't philanthropy — it's a bet that vertical integration (phone + car + robot + AI) creates a moat no specialist can replicate.
  6. The open-source play: publishing MiMo-Embodied lets the world improve your car's brain for free. This is the Android strategy applied to autonomous driving.
  7. The AI metallurgy angle: "AI didn't just drive the car — it invented the metal." 10 million alloy formulas simulated to find one that skips heat treatment entirely. This is AlphaFold for car manufacturing. A PhD student tests 100 compositions in 3 years; Xiaomi's AI tested 10 million and got a better answer.
  8. The compounding loop: AI designs the metal → metal eliminates a factory step → factory gets faster → car gets cheaper → cheaper car sells more → more revenue funds more AI. This is the flywheel nobody's talking about.

Questions Worth Asking on the Show

  • Is this the end of "dumb chassis"? Will buyers expect AI-controlled suspension as a baseline within 5 years?
  • Can Western OEMs catch up on vertical integration, or does the phone-company DNA (fast iteration, software-first) give Xiaomi a permanent cultural advantage?
  • What happens when this car hits European roads with German driving norms — does the China-trained model transfer?
  • If the model is open-source and SOTA, why aren't other Chinese automakers just... using it?
  • At what point does a car with 700 TOPS of AI compute and robotics intelligence stop being a car and start being a robot with seats?

Pitfalls to Call Out

  • Don't oversell: the Electrek first-drive literally couldn't test the AI systems. Claims ≠ verified performance.
  • The "phone company makes cars" narrative undersells that Xiaomi has 3,000+ automotive R&D engineers and Lei Jun committed $10B over 10 years to the automotive division.
  • Benchmarks ≠ real-world. SOTA on 29 benchmarks is impressive, but driving in Beijing traffic is not a benchmark.
  • The regulatory asymmetry: features legal in China may be years from approval in Europe/US. The car you buy may not be the car you read about.
  • Don't conflate "Dragon Chassis" with the fully active suspension pre-research tech. The Dragon Chassis ships now (air springs + CDC). The 4.6kW/wheel active system is future tech.

Connecting Tissue (Episode-Level)

Thread linking to Memory + Multi-LLM deep dive: The Dragon Chassis demonstrates the same principle as AI agent memory — the value isn't in any single component but in unified context. Just as coding agents need shared memory across tools, Xiaomi's chassis needs shared intelligence across driving, suspension, traction, and prediction. Siloed systems (separate chips for ADAS and chassis) are the automotive equivalent of a bloated CLAUDE.md that nothing else reads.

Recurring metaphor: "One brain, many bodies." MiMo-Embodied runs robots and cars. The harness (Cursor's thesis from the Top 5) runs multiple models. The Dragon Chassis runs driving and chassis on one chip. The pattern of the week: unification beats specialization.

Contrarian take: Xiaomi's approach might actually be too unified. A bug in the driving model that also controls your suspension is a safety risk that separate systems avoid. The redundancy of having different brains for different functions isn't just legacy — it's a safety architecture. Mercedes separates these systems by design, not by accident.

Optimistic take: If MiMo-Embodied actually works as described — and the benchmark results plus the open-source release suggest Xiaomi is confident enough to let people look — this is the strongest signal yet that embodied AI is ready for production. Not in a $200K robot. Not in a research lab. In a $32K sedan selling 15,000 units in half an hour.

Derivative Content Ideas

  • Twitter/X thread: "A phone company just built the smartest car chassis on Earth. Here's how the AI actually works (thread)" — 10-12 tweets, architecture diagram as image
  • LinkedIn post: "The $32K car with a robot brain" — frame around what this means for automotive industry talent (AI engineers > mechanical engineers)
  • YouTube short: "The Dragon Chassis explained in 60 seconds" — visual: show the 5-layer AI stack, end with price comparison to Mercedes
  • Newsletter writeup: Deep technical comparison of XLA vs Tesla FSD vs Mercedes Drive Pilot, with implications for 2027 European launches
  • Instagram carousel: Sensor placement diagram, compute comparison bar chart, price comparison

Sources