Cowork Is Now Just Claude — and Docs and Slides Arrive With It
Eight months after launching Cowork as a research preview, Anthropic retired it as a separate product. Announced September 16: the Chat-versus-Cowork toggle goes away, and Claude decides on its own when a conversation becomes a longer-running project, pulling in the context, skills and connectors you already have. Anthropic’s stated reason is the one users kept giving them — “the frustrating part was deciding where a task belonged.”
Shipping alongside: Claude Docs and Claude Slides, both in beta, plus Claude Design working inside any conversation. These are web-native working documents, not files dumped at the end of a chat — they live at one shareable link, stay editable from a phone, and support simultaneous editing and comments. Slides download as PowerPoint or PDF; Docs export to Microsoft Word or Google Docs, with more destinations planned.
The rollout details matter for anyone planning around this. Pro and Max first, on web, desktop and mobile over the coming weeks, with Team and Free to follow. Nothing to turn on. Existing Cowork chats, projects, artifacts, connectors and skills are all retained, and Enterprise admins get at least 30 days notice before anything changes. Claude Code stays separate — Anthropic is explicit that it remains the product for developers working in terminals and IDEs against codebases.
Anthropic’s Policy Chief: No Honor Code, No Self-Grading
At POLITICO’s Decoded Summit in Washington on September 16, Anthropic head of public policy Sarah Heck made the company’s regulatory ask about as plain as it gets: “I don’t think that there’s a world where you do safety and people are accepting of AI companies just doing it on honor code. We can’t be checking our own homework.”
That is a direct rejection of the self-regulation posture congressional Republicans have relied on for catastrophic-risk questions, and it arrives four days after Amodei’s pacing essay. Heck also said Anthropic is in contact with the White House on a daily basis and has been working with Congress all year — which reframes the essay as the public half of a lobbying campaign that has been running quietly for months.
She paired it with the other half of Anthropic’s position: export controls on AI chip sales to China, and the argument that the US needs to keep its lead. That is the same bundle Beijing objected to this week, and it is the bundle that makes this hard to read as pure safety advocacy. Slow the frontier, except against the competitor you want restricted.
The King Gets a Meeting the Regulators Haven’t
King Charles III convened AI leaders at Dumfries House in Ayrshire, Scotland, today — roughly thirty executives, policymakers and ethics researchers. Expected in the room: Nvidia CEO Jensen Huang, Google DeepMind co-founder Demis Hassabis, OpenAI CFO Sarah Friar, UK AI minister Kanishka Narayan, representatives from Anthropic, and Paolo Benanti, the Vatican’s AI adviser.
The stated agenda is how AI can serve society and the natural world, with discussion of a shared set of development principles. The King’s framing, per advance reporting, is that the technology should stay “firmly in the service of humanity, community and the natural world.”
Note the shape of this. A shared set of principles, arrived at privately among the same handful of labs, is structurally the thing Bloomberg raised antitrust questions about two days ago — only with a monarch in the room instead of a mediator. It is also, pointedly, not a hearing. The week’s pattern is holding: the coordination keeps happening in venues where nobody has to testify.
Anthropic Opens Singapore in October, Hires OpenAI’s Regional Lead to Run It
Announced September 16: Anthropic will open a Singapore office in October, its first in Southeast Asia and fifth in Asia-Pacific after Tokyo, Seoul, Bengaluru and Sydney. Running it: Dale Finlay, who arrives after roughly six months as OpenAI’s regional go-to-market head.
The site selection is data-driven rather than sentimental. On Anthropic’s own Economic Index, Singapore ranks 2nd out of 121 countries in Claude usage per capita, behind only Australia. International managing director Chris Ciauri called it “a natural next step.”
Finlay’s line is the one worth stealing if you sell AI into enterprises: “ambition isn’t the blocker to AI adoption — trust is. Once that trust exists, the real work is embedding AI into how the business actually operates.” For context on the competitive temperature: OpenAI put its Asia HQ in Singapore in late 2024 and is reportedly negotiating another 100,000 square feet there.
2.1.274: The MCP Transport Release
Claude Code 2.1.274 shipped today, and it is unusually heavy on plumbing. If you run custom MCP servers, three fixes land directly on you: HTTP servers that only speak legacy HTTP+SSE were failing to connect when they answered the first request with a 422 or other 4xx — fixed. Streamable HTTP tool calls were timing out around five minutes even when you had set a longer per-server timeout — fixed. And MCP prompts and resources weren’t refreshing when a server sent list-changed notifications without declaring listChanged — also fixed.
The bigger structural change: Bedrock, Vertex, Foundry and telemetry-disabled installs now default to the v2 MCP client with MCP 2026-07-28 negotiation against direct HTTP servers, matching what everyone else already had. Opt out with MCP_SDK_GENERATION=v1 or MCP_PROTOCOL_NEGOTIATION=legacy if something breaks. Also new: CLAUDE_CODE_MCP_STARTUP_WAIT_MS to bound how long a first non-interactive turn waits on still-connecting servers (0 means don’t wait), and a visible warning when memory usage goes critical, with steps to free memory or restart safely.
Two quality-of-life wins worth calling out. Sessions stuck in an endless retry loop on “unexpected tool_use_id” 400 errors now self-heal where possible, and where they can’t, you get a clear error with a /rewind hint instead of an infinite spin. And /code-review stopped spawning a swarm of review subagents for models without tuned settings, using leaner inline prompts instead — faster and considerably cheaper.
Three Security Fixes in 2.1.274 You Should Read Before Upgrading Anything Else
Buried in a very long changelog, three items are security-relevant and worth surfacing on their own. First: MCP connection errors and the MCP login tool’s description were displaying secrets resolved from ${VAR} placeholders in MCP configs. If you have been pasting Claude Code MCP errors into tickets or Slack, that is worth a look back.
Second and third are both Bash permission gaps. Commands that loop over or assign certain special shell variables were not triggering a permission prompt — they now ask. And worktree-isolated sessions were accepting Bash commands with certain nested shell expansions — those are now refused outright. Read together with last release’s subshell-hiding-rm fix, that is two consecutive releases hardening the same surface: what the permission checker can and cannot statically understand about a shell command.
The transferable lesson for anyone running agents with shell access: the permission boundary is only as good as the parser behind it, and shell syntax is adversarially easy to make unparseable. If your own agent tooling gates on “does this command look dangerous,” assume the same class of bypass exists in it.
The Auditors Anthropic Promised to Embed Would Like the Terms in Writing
Amodei’s essay committed Anthropic to giving evaluation groups like METR and Redwood Research unprecedented internal access, and Sam Altman matched the pledge. TechCrunch went and asked the evaluators. The answer, as of September 16: neither company has said which evaluators, when, how many, what systems they could access, or what they would be allowed to publish — “despite repeated questions.”
The people being offered the job are not enthusiastic. Adam Gleave of FAR.AI says his firm has already walked away from contracts with frontier developers that wanted too much control over the process. Apollo Research’s Alexander Meinke says the default corporate behavior on training-time misalignment is to neither check carefully nor report truthfully. Palisade Research’s John Steidley reached for the Dieselgate comparison — benchmark-gaming as emissions-test-gaming.
The precedent supports the skepticism. OpenAI gave METR and Redwood roughly a week on-premises for the Hugging Face incident, and gave Apollo three days to test GPT-6 Astra. Meanwhile Meta, SpaceXAI and Google DeepMind have committed to nothing. A pledge whose every operative term is still undefined is not yet a commitment, and the groups who would have to make it real are saying so out loud.
Most of the Lawmakers Regulating AI Have Never Used It
Axios interviewed more than two dozen lawmakers and governors and found that most don’t use AI at all. Sen. Roger Wicker (R-Miss.), 75, who chairs Armed Services, on whether he uses it: “Heavens no.” Rep. Frank Pallone (D-N.J.), 74, ranking member on House Energy and Commerce: “Personally? I don’t think I’ve ever used it” — his office later adding, “You don’t need to do drugs like heroin to know they should be regulated.”
The exceptions are instructive. Rep. Don Beyer (D-Va.), 76, went back to school for a master’s in machine learning and told Axios he prefers Claude. Sen. John Hickenlooper (D-Colo.), 74, on learning his colleagues abstain: “Are you serious? I use it for everything!”
Set that against the same day’s polling. AP-NORC/EPIC found 53% of Americans extremely or very concerned about AI’s environmental impact, up more than ten points year over year, and roughly eight in ten at least somewhat worried about data centers’ effect on power prices, outages or water. A POLITICO poll the same day found more than six in ten see at least moderate risk that AI destroys humanity. Over 100 House Democrats asked Speaker Mike Johnson to cancel recess and vote on AI safeguards. He signaled no.
Anthropic Just Deleted the Word “Agent” From Its Own Interface
Read the Cowork merge as a product decision and it is a sensible bit of UX hygiene: people couldn’t tell which tab to use, so the tabs went away. Read it as a strategy and it is the most aggressive thing Anthropic has shipped this year. For eight months, handing Claude a job that runs for an hour, touches your files and uses your connectors required a deliberate act — you clicked into a different mode. As of yesterday, Claude decides. The agentic path is no longer a door you walk through. It is the default, and the user is not asked.
Anthropic kept the safety rails and they are real — permission-before-action stays on by default, you choose the autonomy level, Enterprise admins get 30 days. But notice what changed anyway. The friction that used to be the consent signal is gone. When Cowork was a toggle, choosing it meant you understood you were delegating. Now the same capability arrives because your phrasing sounded like a project. That is better product design and a meaningfully different consent model, and both things are true at once.
The timing makes it sharper. In one 48-hour window, Anthropic told Washington it cannot be trusted to check its own homework, sent representatives to help draft “shared principles” with a monarch, watched the evaluators it promised to embed publicly ask for the terms in writing — and removed the last interface boundary between chatting with a model and turning one loose on your files. None of those contradict each other, exactly. Anthropic’s whole thesis is that the frontier arrives regardless and the question is who is holding it. But the company is asking for external oversight of a system it is simultaneously making more autonomous by default, and that is a hard posture to hold for long.
The commercial read is simpler and probably more predictive. Docs and Slides are not features, they are a landgrab. Web-native, one shareable link, simultaneous editing, comments, export to Word and PowerPoint. That is a description of Google Workspace and Microsoft 365 with the authorship inverted — the document doesn’t start empty and wait for you, it starts finished and waits for your edit. VentureBeat correctly connects this back to the “SaaSpocalypse” nerves of last summer, and this release is the clearest statement yet of where Anthropic thinks the value sits. Not in the answer. In the artifact, and in owning the surface where the artifact lives. Everything else this week — the policy asks, the summits, the auditor pledges — is happening while that quietly ships.