“When AI Builds Itself”: Anthropic Wants a Verifiable Pause Button
The Anthropic Institute published “When AI builds itself” on June 4, and it’s the most consequential thing the company has put out this quarter. The headline number: more than 80% of the code merged into Anthropic’s production codebase is now authored by Claude, up from low single digits when Claude Code launched in early 2025 — and the typical engineer ships roughly eight times the daily code volume of 2024. The paper maps three scenarios toward recursive self-improvement, where AI begins designing and training its own successor.
What Anthropic asks for is unusual: not a unilateral halt, but a verifiable, multilateral pause mechanism — multiple frontier labs in multiple countries able to agree to stop under the same conditions and confirm the others actually did. A solo pause, the company argues, just changes who leads. Anthropic says it isn’t at full recursive self-improvement yet, but that the world is closer than most institutions are prepared for. Coming the same week its S-1 sits in SEC review, the timing is impossible to ignore.
Claude Code Limits Double — Underwritten by a SpaceX Compute Deal
Anthropic doubled Claude Code’s 5-hour rate limits for every Pro, Max, Team, and seat-based Enterprise account, and killed peak-hour throttling on Pro and Max. On the API side, Opus rate limits went up considerably — third-party coverage cites moves as large as a 1500% jump in max input tokens per minute for some Opus tiers. All of it is effective now.
The capacity behind the generosity is the real story: Anthropic signed an agreement with SpaceX to use the full compute capacity of its Colossus 1 data center — more than 300 megawatts and over 220,000 NVIDIA GPUs — coming online within the month. That’s the supply side of the same flywheel the pause paper describes: more compute feeds bigger limits, which feed the agentic workloads that are now writing most of the company’s code.
Usage Limits Reset, the Subagent Bug Gets Pinned on Opus 4.8 — and “Workflow” Becomes “Ultracode”
Anthropic reset the 5-hour and weekly windows for all Pro and Max users after a rough stretch of runaway token burn. The technical explanation lands the blame not on dynamic workflows but on a flaw in how requests from the Opus 4.8 model were handled — a bug that caused some sessions to spawn excessive parallel subagents and chew through quota. The fix is in, and limits were raised to accommodate higher-effort runs.
One ergonomics change worth knowing: the dynamic-workflow trigger keyword was renamed from workflow to ultracode, so the bare word “workflow” no longer kicks off a run (asking in your own words still works). A related fix stops /effort ultracode from blaming the workflows setting when a model can’t run xhigh — ultracode simply isn’t offered on models that don’t support it now.
June 15 Cutover Is 7 Days Out — Retired Model IDs Start Failing Next Monday
One week remains until June 15, when Claude Sonnet 4 (claude-sonnet-4-20250514) and Claude Opus 4 (claude-opus-4-20250514) retire on the Claude API. Anything still calling those strings begins failing that day. The same date splits subscription usage: programmatic calls via Pro and Max plans move to a separate monthly credit pool, so any headless pipeline riding a subscription needs a plan before the meter changes.
The weekend checklist is unchanged and worth a final pass: grep your codebase for pinned model IDs, repoint to current models (Opus 4.8 is the frontier), claim your agent credit from the account-email instructions, and cap anything running unattended. If you’re on a recent Claude Code build, name a current model as fallbackModel so a stray retired ID degrades gracefully instead of hard-failing.
Project Glasswing Scales to ~150 Orgs — and Claude Security Has Patched 2,100+ Bugs
Anthropic extended Project Glasswing — its initiative to secure the world’s most critical software with the unreleased Claude Mythos Preview model — to roughly 150 new organizations across more than 15 countries. The expansion deliberately covers power, water, healthcare, communications, and hardware, sectors the initial cohort under-represented. Mythos Preview has already surfaced thousands of high-severity vulnerabilities, including issues in every major operating system and browser.
For teams who can’t touch a frontier-grade offensive model, the practical handoff is Claude Security: a product built on public models like Opus 4.8 that scans codebases and suggests patches. Anthropic says it has been used to patch more than 2,100 vulnerabilities in three weeks. For aerospace, defense, and industrial buyers, the timeline that matters isn’t Mythos’s release — it’s the verification program that gates who gets the offensive tooling.
Claude for Small Business Puts Ready-to-Run Workflows Inside the Tools SMBs Already Use
Anthropic launched Claude for Small Business, a package of connectors and pre-built workflows that live inside the apps small teams already run: Intuit QuickBooks, PayPal, HubSpot, Canva, DocuSign, Google Workspace, and Microsoft 365. The pitch is less “chat with an AI” and more “turn the boring cross-app chores into one-click recipes” — invoice chasing, lead triage, content drafting, contract review.
It’s the same connector-and-agent architecture the enterprise side has been building, repackaged for owners who don’t have a platform team. For anyone watching the moat thesis, this is the demand side of the ecosystem story: the more an owner’s accounting, payments, and CRM all speak Claude, the harder it is to swap Claude out for a point tool.
The Pause Paper Is an IPO Document Too
Read “When AI builds itself” next to the S-1 and the tension is the whole story. The same fact — Claude writing 80%+ of Anthropic’s code at roughly human parity — is simultaneously the bull case (a model that accelerates its own development feeds the run-rate that justifies a $965B round) and the risk case (a system approaching the ability to design its successor). Anthropic is, remarkably, the one publishing both halves.
The cynical reading is that calling for a verifiable global pause is free if you believe no rival will agree to it — you get the safety halo without slowing down. The more charitable reading is that a company about to face public-market discipline is putting its biggest risk in writing before an underwriter makes it. Either way, the ask is structurally smart: a multilateral, verifiable mechanism is the only kind that doesn’t simply hand the lead to whoever doesn’t stop. Investors will have to price a company that says its core asset might need a brake.
Two Outages in a Week, Still No Written RCA
The reliability column keeps writing itself. After a major June 2 disruption traced to Claude Code’s subagent system spiraling into a runaway loop, Claude took another roughly two-hour hit on June 5 from elevated error rates across Opus 4.7 and 4.8 — and Anthropic is investigating, but has not confirmed, claims that the earlier event may have exposed customer data. Engineers have so far attributed the outages to infrastructure rather than a breach.
The gap that matters: there is still no published post-incident report for June 2. A company pitching itself to public markets as critical AI infrastructure — and simultaneously asking the world to trust a verifiable pause regime — is the same company that owes its customers a written root-cause analysis. Trust in the pause proposal and trust in the status page are, increasingly, the same currency.