LLM Weekly — W31: Opus 5 ships and Anthropic deletes 80% of Claude Code's prompt

The biggest LLM story this week isn't a benchmark — it's a deletion. Anthropic shipped Claude Opus 5 and then cut more than 80% of Claude Code's system prompt, because the model no longer needs the instructions to behave itself.

Anthropic ships Claude Opus 5 — and deletes 80% of Claude Code's system prompt to fit it. Opus 5 lands as Anthropic's new flagship at near-frontier intelligence, and the ship note that matters most is the one about the prompt: over 80% of Claude Code's system prompt is gone because the model doesn't need it anymore. The shrinking prompt is the real signal. Capability is being absorbed into the weights, so the instruction-engineering that defined the last two years is turning from moat into liability. If you hard-coded workarounds for older models, migrating up now means deleting prompt, not adding it. Via AI Search and The AI Opportunities.

Google is already pretraining Gemini 4 — with Gemini 3.5 Pro still in testing. Google acknowledged that pretraining on Gemini 4 is underway even though 3.5 Pro hasn't shipped, and a "Gemini 3.6" reference surfaced separately in the same week. The version numbers matter less than the tempo: Google is running two-plus generations through the pipeline at once, a cadence only a hyperscaler with captive TPUs can sustain. The frontier race is now measured in overlapping training runs, not launches. Via The Creators' AI and AI Search.

Model distillation becomes trade policy: Treasury threatens sanctions over Moonshot and Anthropic's Fable. The US Treasury reportedly threatened sanctions over allegations that China's Moonshot AI distilled Anthropic's Fable model to train its own systems, while Washington and Beijing separately set formal AI talks for September. Distillation just went from engineering technique to export-controlled-IP dispute. For anyone with a China footprint, model provenance and training-data lineage are now a compliance surface, not a research footnote — and Moonshot's own model availability could be the thing at stake. Via The Creators' AI.

Microsoft's MAI-Cyber-1-Flash: a 5B-active model that scores 95.95% on CyberGym. Microsoft's in-house AI group shipped a security-tuned model with roughly 5B active parameters that beats far larger generalists on the CyberGym benchmark. A small, cheap, specialised model topping a security eval is the pattern worth tracking: the economics keep favouring tuned vertical models over frontier general ones for well-defined tasks, and defensive tooling is exactly the kind of narrow, high-value target that rewards it. Via MarkTechPost.

Researchers escaped the sandbox in Cursor, Codex, Gemini CLI, and Antigravity. Security researchers demonstrated sandbox escapes across four major AI coding tools, meaning agent code that was supposed to be contained could reach the host. "The agent runs in a sandbox" is no longer a security control you can assume. If you're running autonomous coding agents on developer machines or CI runners, that box is now the trust boundary — and it leaks. Via The Creators' AI.

The thread running through the week: as capability moves into the weights, the work moves out of the prompt and into everything around the model — provenance, sandboxing, and which lab can afford the compliance. Watch whether the next frontier launch keeps up the deletion trend, or whether Opus 5's prompt diet stays a one-off. If shrinking prompts become the norm, the moat you spent 2025 building may be the first thing your next model migration throws away.

---

This post is also published on our Substack newsletter at edge-ai.forum. Subscribe for the weekly roundup direct to your inbox — fresh AI news, executive context, and devices + robotics every Friday morning.