The Model War Just Became a Price War

Google made Gemini 4 Carbon generally available yesterday, November 10. The day before, Anthropic shipped Claude Fable 5.2 at $8 per million input tokens and $40 per million output tokens. Carbon’s introductory API price is $2 and $10. If you automate real work with agents, that gap hits the loop. The model war just became a price war.

What Carbon is

Google calls Carbon its strongest model, with frontier performance in cybersecurity and in long-horizon agentic work across knowledge work, software engineering, and scientific reasoning. The post from @Google draws the line with Gemini 4 Argon in one sentence: “If Argon is great for complex tasks, Carbon is the model you bring in for your hardest problems.” The blog fills that in. Argon stays the choice for most software engineering and knowledge work. Carbon is for tasks that test the limit, like multi-step coding and debugging and open-ended research, and it spends more tokens than Argon to look harder, reason longer, and run verification loops.

“Strongest” is the headline. A model that spends more tokens on purpose is a different bill. Carbon is on the API in Google AI Studio and Gemini for business, for paying developers and enterprises, at the same introductory price as Argon. It is also rolling out to Google AI Ultra subscribers in Google Antigravity, the Gemini app, and AI Mode in Search. I have not run it. This is a read of yesterday’s posts.

The rate card

Fable 5.2 is $8 in and $40 out. Artificial Analysis called that a 20% cut from Fable 5.1, and still put Fable’s max-effort cost per task at $26.5. That is about 2.6× Claude Opus 5.5 at max effort ($10.0) and about 2× Fable 5.1 at max ($12.99). A lower sticker is still an expensive agent when the effort stays pegged at the top.

Carbon’s introductory price is $2 per million input tokens and $10 per million output tokens, the same rate as Argon. Google’s footnote says the introductory period expires on November 30, 2026. After that, public Carbon is $4 and $20. The footnote limits that later rate to the Carbon model intended for public access. At the introductory rate, Carbon’s list price is a quarter of Fable’s, on input and on output. After November 30 it is half. The token price doubles.

List price meters the token. A finished task also pays for the plan, the tool calls, the retry, and the check. Google says that on the benches below, Carbon’s cost per task is more than 2× lower than the next-best external model. That claim is theirs. Carbon also uses more tokens than Argon to be thorough, and /swarm puts more agents on one prompt, so a cheaper token can fund a second pass or disappear into extra volume. I would write both rates on the trace before a workflow becomes the morning default. A loop that only feels worth it at $2 / $10 is a trial with a date on it.

What /swarm actually does

Google is previewing /swarm in Antigravity. One prompt becomes a coordinated team of agents you can track in real time. Put Carbon on the roles that need the deepest reasoning: architectural design, coordination, and auditing. Pair it with faster, lighter models across parallel isolated workspaces, and with adversarial verifiers that test every change.

That is the shape I want when a job is too big for one pass. A strong model on the plan and the audit. Faster models on the isolated chunks. A verifier whose job is to test the change. The trace is what I would check first: which agent did what, and which verifier failed.

The demos are large on purpose. A team of Carbon agents turned historical photos and engineering plates into an interactive Three.js exhibition of the Eiffel Tower’s 1887–1889 construction, from a single prompt. Another run took seven public City Council planning PDFs, 206 extracted drawing views, and built an interactive reconstruction of Google’s Platform 37 office across 10 overnight loops. I do not have an Eiffel Tower to model. I have piles of files and a draft I would otherwise assemble by hand. Same shape, smaller scale, and I have not watched it run. Each extra agent is another meter, at whichever rate card you handed it.

The benches point different ways

The published numbers disagree, and some of them are different tests at different effort settings. That is a reason to keep a router.

Artificial Analysis, on November 9, put Fable 5.2 at 65 on the Intelligence Index v5 at max effort: five points ahead of Gemini 4 Carbon at high effort (60), and six ahead of Opus 5.5 at max (59). Max and high are different settings. The same note has Fable at 62 on high and xhigh effort, ahead of every other model at any effort on that index. Inside the same writeup, Carbon is ahead of Fable on Terminal-Bench 4.0, 65% to 60%, and narrowly ahead on AutomationBench-AA, 78% to 76%. Fable leads the index. On those two, Carbon is ahead of Fable.

Anthropic’s table shows Fable ahead on FrontierCode 2.0, 49.1% at Xhigh effort against Carbon at 41.1% on Medium effort, and ahead on FrontierSWE v2, 69.9% to 61.2%. The same table has Carbon ahead on AutomationBench, 53.9% to 51.4%. The FrontierCode efforts do not match, so that lead stays labeled as Xhigh versus Medium. On AutomationBench, in Anthropic’s own table, Carbon is in front.

Google’s launch post says Carbon ranks first on CursorBench 4.0, Perplexity’s WANDR, and Zapier’s AutomationBench, at more than 2× lower cost per task than the next-best external model. The post gives the rank and the cost claim, and it does not give the scores. CursorBench is the one I would watch for an agent in an editor. Zapier’s AutomationBench is the one I would watch for business workflows. Those names are Google’s, and they are separate from Anthropic’s AutomationBench and from Artificial Analysis’s AutomationBench-AA.

Anthropic already describes a router. For most coding and knowledge tasks, start with Opus 5.5, move up to Fable for ambitious open-ended work, or down to Sonnet 5.5 or Haiku 5.5 when the job is smaller. Fable can sit as an Advisor the smaller models consult, or as an Orchestrator that hands simpler tasks down. Google draws the same kind of line: Argon for most of the work, Carbon for the hard problem.

What I would actually change

Three habits. I have not moved a workflow onto Carbon yet.

Route by the task. A long verification loop, or a workflow bench where Carbon is the model in front, is a Carbon try, with the token count visible, because the model is built to spend tokens. A hard coding job in the neighborhood of Anthropic’s FrontierCode and FrontierSWE results is a Fable try, and $8 / $40 is what you pay with the effort left high. Ordinary steps stay on the daily driver. Anthropic’s name for that driver is Opus. Google’s is Argon.

Keep more than one model on the hook. /swarm packages a split the work already wants: a strong model for the plan and the audit, a faster model for the parallel pieces, a verifier with its own job. A small version of that split does not have to wait on Antigravity. One model for every prompt puts a frontier rate on a rename-and-file task.

Price the later rate before the introductory rate becomes the assumption. Run one real pile of work, multiply the tokens by $4 / $20, and see if you still want the loop after November 30. If it only works at $2 / $10, it stays a trial.

The models will move again. The piece I can keep is a router, a trace, and a rate card I re-read when November 30 shows up.