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

This edition combines the three newsletters we published separately this week (LLM Weekly, Devices & Robotics and Executive Roundup). Their text is unchanged.

The week in models

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.

Devices & robotics

It was a slow week on the show floor — nothing you could pre-order. The action was one layer down, in the capital and the parameter counts.

1X's pitch deck leaks, and the humanoid thesis is "automate physical labour." A widely-circulated breakdown of 1X's fundraising deck surfaced this week, framing the humanoid company's bet in a single line — "for 200 years technology automated cognition; now it automates physical labour" — with OpenAI among its backers. Read the deck as a fundraising artifact, not a shipping product: it's a story about capital conviction, not a robot you can buy today. But the money moving into embodied AI is real, and a frontier lab writing checks into a humanoid company tells you where it thinks the next platform lives — off the screen and into the room. Via Product Market Fit.

Microsoft's MAI-Cyber-1-Flash: 5B active params, 95.95% on CyberGym. Microsoft's in-house AI group shipped a security-tuned model with roughly 5B active parameters that scores 95.95% on the CyberGym benchmark — beating far larger generalist models on the eval. For a devices audience, the parameter count is the headline: 5B active is small enough to live at the edge, not just in a datacenter. The thing that makes on-device inference real isn't a bigger frontier model — it's a small, cheap, task-tuned one that fits on hardware you already own and still wins on a narrow job. Via MarkTechPost.

The quiet through-line: capability is getting small enough to leave the cloud. Two things happened in parallel. Microsoft's tiny cyber model topped a benchmark, and across the frontier, models kept needing less scaffolding to behave — Anthropic deleted more than 80% of Claude Code's system prompt for its new flagship. Inference Both point the same way for hardware: as capability compresses into smaller weights and models need less hand-holding, the case for running inference locally — on an NPU, a phone, a robot's onboard compute — gets stronger every quarter. The edge-AI story of 2026 isn't a new chip; it's models finally small enough to meet the silicon that's already shipping. Via AI Search.

What to watch next week: whether any of the humanoid contenders — 1X, Figure, Apptronik — turns capital into a deployment number: a count of robots actually on a floor, not a line on a slide. That's the metric that separates the embodied-AI narrative from the embodied-AI business.

What it means for leaders

Two curves split this week. Capability kept climbing — Anthropic shipped Claude Opus 5 and ChatGPT crept toward a billion weekly users — while the money and politics beneath it cracked. The same people who built the "AGI is coming" trade are now either losing money on it or asking Washington to slow it down.

If you're a CEO this week...

Your board's question this week isn't about a model — it's about a blow-up. Leopold Aschenbrenner's ~$45B Situational Awareness fund took steep AI losses, and Citadel bought its stock book at a discount: the author of the AGI manifesto himself is now the sharpest proof that being right about the technology and right about the trade are different bets. Expect your largest investor to probe how concentrated your own AI exposure runs. Second signal: the accelerationists want a brake. Sam Altman and Dario Amodei backed a petition asking Washington to "pace" the frontier, while AI labs set a Q2 lobbying record — incumbents inviting rules only they can afford. Meanwhile the revenue is landing on infrastructure: Azure grew 43%. The question to walk in with: are you positioned like the picks-and-shovels layer that's monetising, or the frontier trade that just took losses?

If you're a CIO/CTO this week...

The architecture signal is Anthropic deleting 80%+ of Claude Code's system prompt for Opus 5 — capability is moving into the weights, so migrating to a stronger model now means removing prompt scaffolding, not adding it. Re-baseline your prompts before the old workarounds start hurting output. On security, treat "the agent is sandboxed" as unverified: researchers broke out of the sandbox in Cursor, Codex, Gemini CLI, and Antigravity, and Hugging Face was hacked — your developer machines and weight-pulls are the trust boundary now. On compliance, Treasury's sanctions threat over Moonshot allegedly distilling Anthropic's Fable makes model provenance a real audit surface. Build-vs-buy read: for defined security tasks, buy small and specialised — Microsoft's 5B-active MAI-Cyber-1-Flash hit 95.95% on CyberGym — don't wait on a frontier generalist.

If you lead AI transformation this week...

Your bridge job this week is turning the Opus 5 prompt cut into a practice, not a headline. The technique to teach: subtractive prompting — ablate one instruction block at a time, keep only what breaks when removed. It reframes the prompt-engineer role from adding scaffolding to pruning it, and that skill shift is your change-management item. The deeper pattern cuts across the whole week: as capability commoditises, durable value moves to owning the workflow and the outcome, not model access — the Sequoia "services are the new software" thesis drew a sharp rebuttal, and proof points like a three-person, seven-figure agency run on Claude show where the leverage sits. On governance, agent autonomy now needs oversight you can't assume. The experiment to run this month: take your three most over-engineered prompts, ablate them against Opus 5, and measure whether leaner beats longer.

What all three seats share this week: the capability curve is no longer the hard part — the hard part is capital discipline, model provenance, and workflow ownership layered on top of it. The question every seat should be asking Monday is the same: are we betting on the model, or on the work it does?

---

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.