Anthropic Says Claude Now Leads a Quarter of Its Own R&D
Anthropic did something yesterday it has never done: it published an internal number. An internal review of research and development tasks found Claude was able to “lead” that work roughly 26% of the time — meaning the model completes most of a task end-to-end from a high-level prompt, with humans supervising and providing direction rather than driving. As recently as February, that figure was zero.
The second number is arguably the more interesting one. Anthropic says about 90% of its R&D is now done “in collaboration” with Claude, which the company defines as the model doing large chunks of work under close human direction. So the 26% is not a separate track running in a corner — it is the top slice of a workflow that has almost entirely reorganized itself around the model in about seven months.
Anthropic was careful on the ceiling: it says there is no point today at which Claude operates fully autonomously without a human in the loop. That caveat is doing real work in the announcement, because the thing being described — a model meaningfully building its own successor — is the exact scenario Dario Amodei spent 20,000 words last week arguing the industry needs to slow down for. The company published both. Within six days of each other.
Anthropic Picks Nasdaq, Targets October
Anthropic has selected Nasdaq as the venue for its planned IPO and is targeting an October listing. Reporting puts the potential valuation near $2 trillion, though that figure is explicitly not finalized. For scale: the Series H in May valued the company at $965 billion, and secondary-market activity has since implied somewhere between roughly $1.05T and $1.5T.
The underlying numbers are what make the range arguable rather than absurd. Annualized revenue sits around $65 billion. The offering is expected to raise more than $60 billion, with Goldman Sachs, JPMorgan and Morgan Stanley named as lead underwriters. Anthropic expects a second consecutive profitable quarter — which, in this cohort, is the genuinely unusual part.
If the $2T mark holds, it would clear SpaceX’s $1.75 trillion listing in June and make this the largest tech debut on record. Worth holding onto for the next few weeks: a company about to face quarterly earnings calls is a company whose safety commitments acquire a shareholder audience.
Smart Reports: Claude Reads Your Team’s Transcripts and Writes the Memo
Smart Reports is in beta on Claude Enterprise plans, and it is a more pointed feature than the name suggests. Rather than session counts and spend dashboards, it produces prose and charts covering what kind of work actually got done, what it cost, and where people ran into friction — plus which repeated patterns are worth packaging as shared skills.
Setup is scoped: you pick a team or department, a time range up to the last 28 days, the products to include, focus areas, and any custom questions you want answered. Claude then reads a sample of the transcripts in scope, groups them into workstreams and output types, attaches spend to each group, and writes it up. Ten reports per org per month, free during beta.
The exclusions are the part to read carefully before you promise anyone a rollout. Organizations using customer-managed encryption keys, HIPAA configurations, or Access Transparency cannot use smart reports, and neither can Enterprise orgs running Claude Code with zero data retention. That is a coherent list — the feature works by reading transcripts, so every configuration built to stop anyone reading transcripts is out.
2.1.275 Brings Skill and Plugin Sync — and 2.1.276 Fixes What It Broke
Claude Code 2.1.275 is a genuinely useful release. The headline additions: syncClaudeAiSkills and syncClaudeAiPlugins settings, which pull your claude.ai skills and plugins down into the CLI instead of leaving you to maintain two parallel sets. Alongside them, a --marketplace <source> flag on /plugin install, and a Ctrl+Enter (or Ctrl+X Ctrl+S) send-now keybinding for anyone who has ever fought the newline-versus-submit problem.
The quiet security item: npm-sourced plugin installs now run integrity verification via npm pack --ignore-scripts. That matters more than it reads. Installing a plugin from npm previously meant trusting whatever lifecycle scripts the package declared — and plugin ecosystems are exactly where supply-chain attacks go looking. If you install plugins from npm sources, upgrade for this alone.
Then 2.1.276 arrived within hours, and it exists for one reason: 2.1.275 made every request fail with a 400 'Input tag advisor_20260301' error when ANTHROPIC_BASE_URL pointed at a proxy or gateway. If you route Claude Code through a corporate gateway, LiteLLM, or any self-hosted proxy, skip 275 and go straight to 276. There is nothing else in 276 — it is purely the regression fix.
Workbench Is Now Playground, and the Messages API Learned to Compact
Two platform changes worth a config review. First, Workbench has been renamed Playground in the Claude Console — not just a label swap. Playground now supports every Messages API parameter, ships templates demonstrating code execution and web search, and shows the full SDK request alongside the API response for each run. That last bit turns it from a prompt scratchpad into something you can actually copy working code out of.
Second, and more consequential for anyone running long agent loops: the Messages API can now compact a conversation on demand, in beta behind the compact-2026-09-04 beta header. Compaction has been a client-side problem everyone solved differently and badly. Having it as a server-side primitive means one less hand-rolled summarizer in your stack.
Two smaller items: a new anthropic-workspace-id response header carries the wrkspc_-prefixed ID of the workspace your key or token resolved to — handy for multi-workspace billing attribution and for catching the wrong-key-in-prod class of mistake. And the Compliance API session endpoints are out of beta for Cowork and Claude Code sessions, with the Admin API now in the ant CLI and the Python, TypeScript, C#, Go, Java, PHP and Ruby SDKs under client.beta.organization.
Six Days In, the Cross-Lab Safety Talks Still Have No Written Terms
Dario Amodei published “We Must Pace the Frontier” on September 12, arguing that pacing means not halting training but building in enough time for alignment work, third-party verification and operational rigor to keep up with model capability. The response was fast and near-unanimous: Sam Altman posted that he agreed and would match Anthropic’s first commitment; Elon Musk wrote “Dario is right”; DeepMind leadership aligned publicly within hours.
OpenAI global policy chief Chris Lehane subsequently told reporters the company has been working with Anthropic and Google DeepMind on AI safety for weeks — meaning the talks predate the essay that appeared to prompt them. That is either reassuring or the opposite, depending on whether you think coordination announced after the fact is coordination or choreography.
What still does not exist, six days later: written terms. Not which evaluators, not what access, not what anyone would be permitted to publish. The critique that landed hardest — that the essay was less about safety than about consolidating control — stays live precisely as long as the operative details do not.
Claudeforce Hits Open Beta This Month
Salesforce in Claude — the customer-facing half of the Claudeforce partnership announced August 26 — moves from select pilot customers to open beta in September. The arrangement makes Claude the default reasoning engine across the Salesforce ecosystem, embedded into the Agentforce Atlas Reasoning Engine that powers Agentforce Vibes and Coworker.
The commercial shape is worth restating because it is unusual. Salesforce is spending roughly $300 million on Anthropic tokens in 2026 alone, on top of an existing $300 million equity stake. That is not a vendor relationship with an exit ramp.
The read several analysts have converged on: this is the end of model-agnostic enterprise AI as a default posture. Salesforce is betting the future of work is agent-platform architecture rather than individual chatbot benchmark wins — and if that bet is right, the platform layer is where lock-in accrues, not the model layer.
Zero to 26 in Seven Months Is the Only Number That Matters Today
Strip out the IPO noise and the platform release notes and one fact is carrying the week: in February, Claude led none of Anthropic’s R&D. Today it leads 26% of it. Not assisted. Not autocompleted. Led — end-to-end from a high-level prompt, with a human supervising. That is a seven-month slope, and the honest thing to say about a seven-month slope with one data point of history is that nobody knows its shape, including Anthropic.
The company’s framing is careful and probably sincere: no fully autonomous operation, human always in the loop, 90% collaborative. But notice what the 90% figure actually concedes. It is not a safety boundary, it is a saturation measurement. There is almost no R&D left at Anthropic that doesn’t route through Claude. The remaining question is not whether the model participates — that’s settled — but how much of each task it leads, and that number went from 0 to 26 while the org chart stayed the same size.
Which makes the timing genuinely awkward rather than merely ironic. Six days ago Amodei argued the industry must build in enough time for alignment work and third-party verification to keep up with capability. Yesterday his company disclosed that the capability side of that race is being run increasingly by the thing being verified. Those positions are reconcilable — you can believe recursive R&D is exactly why external evaluators need embedding, and Anthropic clearly does. But the verification half still has no signed terms, no named firms, and no publication rights, while the capability half just posted a number.
The IPO is the pressure the analysis usually misses. A company at roughly $65B annualized revenue heading for an October listing is about to acquire quarterly earnings calls, and a self-improving research loop is a spectacular thing to put in an S-1. The 26% disclosure reads as a safety-transparency gesture and it may well be one. It also reads as a growth story with a compounding curve attached, published four weeks before a roadshow. Both can be true — and if you are trying to forecast what Anthropic does next, assume the number gets updated, and assume it goes up.