Abstract. On September 16, 2026, Anthropic discontinued Claude Cowork as a separate product and folded it into a single, unified Claude. The mode toggle that used to make you choose between "quick chat" and "multi-step agentic project" is gone; starting with Pro and Max plans and rolling out to Team and Free over the following weeks, Claude now makes that call itself, deciding on its own whether a prompt gets a conversational reply or a longer run that searches the web, reads files, runs code, and assembles a finished result. Source: VentureBeat, "Anthropic is killing off Claude Cowork and folding it into Claude chat, launching Claude Docs and Claude Slides," September 2026. Source: Unite.AI, "Anthropic Folds Cowork Into a Single Claude Experience Across Plans," September 2026. Anthropic shipped two new beta tools alongside the merge, Claude Docs and Claude Slides, and was direct about what the new autonomy costs: "longer agentic tasks generally use more than a quick question." That's a second "auto" showing up in the same stack this blog already tracks one of. Claude's new auto decides how much work a request deserves. A cost-aware router's auto decides which model a request deserves. This post lines the two up, because the first one just made the second one more necessary, not less. What actually changed on September 16. Cowork existed as a mode you picked. You opened a chat for a quick question, or you opened Cowork when you had a real task, research, a report, a spreadsheet, a deck, and wanted Claude to work through it with less supervision. Anthropic's own reasoning for killing that split: "People used both, and told us the frustrating part was deciding where a task belonged." Source: Unite.AI, September 2026. The fix removes the decision from the user entirely. Describe what you need, and Claude decides which capabilities to deploy, a permission setting (Manual or Auto) controls how much it's allowed to do on its own, and for genuinely open-ended work it can search, read, run code, and hand back a finished, editable file. Source: VentureBeat, September 2026. Multiple tasks can run at once, and a long-running job keeps working in the cloud after you close the tab. Source: Unite.AI, September 2026. Two new tools ship in beta alongside the merge: Claude Docs, where Claude drafts sections and colleagues edit them live, private by default and exportable to Word or Google Docs, and Claude Slides, which builds a deck you can edit in place and download as PowerPoint or PDF. Claude Design now sits directly in the chat window for graphics instead of living as a separate surface. Source: VentureBeat, September 2026. Existing chats, projects, artifacts, connectors, and skills carry over automatically. There's no rollback once an account migrates. Source: Unite.AI, September 2026. Two autos, one bill: Claude's new auto decides scope, a router's auto decides model. Claude's auto turns a prompt into either a chat reply or a multi-step project that searches, reads files, runs code, and keeps working in the cloud, with no scope dial exposed and Anthropic's own admission that longer tasks cost more. A cost-aware router's auto scores the same request against a quality floor, sends it to the cheapest model that clears the bar, and reports the model and the exact cost back on every response. The line that matters is the one about cost, not the one about features. Buried in the announcement is the only sentence that actually prices the change: longer agentic tasks generally use more than a quick question. That's not a caveat. It's an admission that the thing being removed, a human manually opting into the expensive mode, was also the thing keeping a lid on spend. This blog has already measured the multiplier that "expensive mode" carries: Anthropic's own engineering team found a single agent burns roughly 4x the tokens of a chat turn, and a multi-agent system fanning work out to subagents burns roughly 15x. A deep-research-style loop, which is functionally what "search the web, read files, run code, assemble results" describes, runs on the order of 10x the tokens of a normal chat turn by design, because it's paying for fan-out and synthesis, not just an answer. None of that is new behavior. What's new is who decides to invoke it. Before September 16, a user clicking into Cowork was a legible signal: this task is worth the bigger loop. After September 16, that signal is gone, and the decision moves to a model that has to guess, from the prompt alone, whether "help me put together our Q3 numbers" means a two-paragraph summary or a multi-step research project with a Docs file at the end. Anthropic's own permission setting, Manual versus Auto, is the tell: even the company shipping this knows the autonomous version needs a dial, and shipped one. The dial controls how much Claude is allowed to do on its own. It does not tell you, before the run, what that run is going to cost. The pattern this blog has already named twice. Apple, Perplexity, and Google all shipped routing architecture in the same week back in June, and the conclusion then was that the routing layer stopped being optional the moment three platforms converged on it independently. Anthropic just added itself to that list, one layer up: instead of routing between models, Claude's Auto routes between task scopes, chat turn or agentic project, using the same "let the system decide instead of the user" logic. Meta spent ten months finding out what happens when a company rewards AI usage volume without pricing the outcome, and reversed a formal performance-review policy over it in the same week as this announcement. The mechanism is identical: a system that can expand its own scope, its own token spend, or both, without a corresponding number attached to the decision, will expand until someone downstream notices the bill. Claude Docs and Claude Slides sit on top of the same gap. Structured, formatted output already carries a measurable tax over plain prose, schema tokens on the way in, verbose formatting on the way out, and a live-edited document or a rebuilt slide deck means that generation can happen more than once per task, at whatever length Claude judged the doc needed. A five-slide deck and a forty-slide deck are the same feature from the user's chair. They are not the same number of tokens, and nothing in the new flow surfaces that difference before the file lands. What a request looks like when the scope decision and the model decision are both automated. Claude's Auto answers "how much work does this deserve." It doesn't answer "which model should do that work, and what will it cost." Those are separable questions, and the second one is the one a routing layer already has to answer on every call, agentic or not: import openai client = openai.OpenAI( base_url="https://api.getnadir.com/v1", api_key="ndr_...", ) Whether this call is one turn in a quick chat reply or the fortieth sub-step of an autonomous multi-file project, it still arrives at the endpoint as one request. model="auto" scores it against your quality floor and picks the cheapest model that clears it, on that request, not on a mode chosen once for the whole task. response = client.chat.completions.create( model="auto", messages=[{"role": "user", "content": prompt}], ) Every response reports what actually happened, regardless of which scope decided to make the call in the first place. model_used = response.model cost = response.model_extra["nadir_metadata"]["cost"]["total_cost_usd"] Same call shape whether the caller is a person typing into a chat box or an autonomous loop that decided, on its own, that this prompt deserved a multi-step run. Nadir doesn't need to know which one it was to price the request correctly. That's the point: when the layer above you starts making scope decisions you can't see coming, the layer below needs to keep pricing every individual call regardless, because the alternative is finding out what "longer agentic tasks generally use more" meant on next month's invoice. What to check before this rolls out to your team. Find out which plan you're on. Pro and Max get the unified experience first, Team and Free follow. Source: Unite.AI, September 2026. If your team is on Team or Enterprise, you have a short window before this ships to decide what "Auto" permission setting you want as the default, not after adoption is already underway. Set the permission dial deliberately, don't leave it on the default. Manual versus Auto is the one lever Anthropic gives you over how much scope Claude claims for itself. Treat it the way you'd treat a spend cap, a policy decision, not a setting nobody revisits. Instrument cost per task, not cost per seat. A flat per-seat number was already the wrong metric for measuring AI usage before scope itself became autonomous. Now it's measuring even less: the same seat can trigger a two-cent reply or a two-dollar multi-file project, and a per-seat average hides which one happened. Route what sits underneath the autonomy, since you can't route the autonomy itself. You don't get to decide whether Claude expands a prompt into a project. You can decide which model handles every call inside whatever it decides to run, and that's the decision a quality-floor router makes on your behalf, at the call level, regardless of which scope a product-layer auto chose above it. Anthropic removed a toggle and was honest about the tradeoff in the same sentence: less friction, more tokens on the runs that used to need a deliberate click. That's a reasonable product bet. It also means the manual signal that used to tell finance which requests were expensive is gone, and the only number left that still means anything is the one attached to the individual call, not the mode, not the seat, not the plan. Two autos now sit in the same stack. Only one of them was built to tell you what it cost. Related reading Apple, Perplexity, and Google all shipped routing this month. The pattern is the architecture. Multi-agent AI costs 15x more, and almost nobody routes it. Deep research agents run a 5-to-100x price spread for the same feature, gated by a manual mode switch nobody automated. Meta spent ten months rewarding AI token volume before reversing the policy in September 2026. Structured output schemas and JSON verbosity add a measurable tax most teams have never priced. Sources: VentureBeat, "Anthropic is killing off Claude Cowork and folding it into Claude chat, launching Claude Docs and Claude Slides," September 2026. Unite.AI, "Anthropic Folds Cowork Into a Single Claude Experience Across Plans," September 2026. Quotes attributed to Anthropic ("People used both..." and "longer agentic tasks generally use more than a quick question") as reported by the sources above; Nadir has not independently verified Anthropic's internal usage data.