Google Opened Gemini 4 Carbon at a Quarter of Fable 5.2's Price, a Day After Anthropic's Cut. On December 1, Carbon Costs What Opus 5.5 Does.
Two frontier models went on sale a day apart this week. On Monday, November 9, Anthropic released Claude Fable 5.2 and Claude Mythos 5.2 and cut Fable's API price to $8 per million input tokens and $40 per million output, from $10 and $50. On Tuesday, November 10, Google opened Gemini 4 Carbon, which it calls its strongest model, to paying developers, businesses and Google AI Ultra subscribers.
Google's price is $2 and $10. That is a quarter of Fable 5.2 on both lines. It also does not last: Google's pricing page lists those rates through November 30 and $4 and $20 starting December 1. At that point Carbon costs exactly what Claude Opus 5.5 does, not what Fable does.
We already covered the Fable cut as a side note to Anthropic's IPO range yesterday. What changed is that the comparison now has a second column, and it was written by Google.
What Opened, and Where
Google announced Carbon on Wednesday, October 28, and gave it first to cyber defenders in its Fairwind Program. Tuesday's rollout came 13 days later. Carbon is now on the paid tier of the Gemini API through Google AI Studio and Android Studio, and in Gemini for business. It is not on the free tier.
For consumers, it is rolling out to Google AI Ultra subscribers in the Gemini app, AI Mode in Search and Antigravity, Google's coding tool. Android Authority notes that Ultra starts at $100 a month in the US, that AI Mode access carries daily limits and an 18-plus age requirement, and that Google gave no timeline for cheaper plans. Google also previewed /swarm in Antigravity, which splits one prompt across a team of agents, with Carbon handling design and coordination while lighter models work in parallel.
Google itself does not position Carbon as the default. Its post says Gemini 4 Argon remains the right pick for most software engineering and knowledge work, and that Carbon spends more tokens reasoning and checking its own work on harder tasks. The Next Web makes the practical point: the same task can cost more on Carbon than on Argon at the same per-token price.
The Price Lines, Side by Side
The table uses each company's published standard API rates. The last column is TensorFeed's illustration: one uncached job with 2 million input tokens and 400,000 output tokens. Real token counts differ by model, so read it as a price ratio, not a task cost.
| Model | Input / 1M | Output / 1M | Cached input / 1M | 2M in, 0.4M out |
|---|---|---|---|---|
| Gemini 4 Carbon (through Nov 30) | $2 | $10 | $0.10 | $8 |
| Gemini 4 Carbon (from Dec 1) | $4 | $20 | $0.20 | $16 |
| Claude Opus 5.5 | $4 | $20 | $0.20 | $16 |
| Claude Fable 5.2 | $8 | $40 | $0.20 | $32 |
| Claude Fable 5.1 | $10 | $50 | $0.25 | $40 |
| GPT-6 Astra | $10 | $50 | $1.00 | $40 |
Sources: Google's Gemini API pricing, Anthropic's Claude pricing and OpenAI's API pricing, standard tier (OpenAI's short-context rate for Astra). Gemini 4 Argon has the same rates and the same December 1 step as Carbon.
One line in that table is easy to miss. From December 1, Carbon, Opus 5.5 and Fable 5.2 all charge $0.20 per million cached input tokens. Anthropic prices Fable's cache reads at 2.5 percent of base input, against 5 percent for Opus 5.5, so on that line the gap disappears. On a long agent run that mostly rereads its own context, the blended gap between Fable and the cheaper models falls below the 2 times the sticker shows, though output tokens still cost twice as much. Fast mode is a separate story: Fable 5.2's optional fast tier costs $16 and $80.
Anthropic's Own Table, Read for Carbon
Anthropic put Carbon in its Fable 5.2 launch table, a day before Google opened it to the public. Eight rows include a Carbon figure. Fable 5.2 leads on seven. Carbon leads on one, AutomationBench, the Zapier business workflow test, 53.9 percent to 51.4, a result Let's Data Science also reported from the release. All numbers below are as Anthropic published them.
| Benchmark | Fable 5.2 | Opus 5.5 | GPT-6 Astra | Gemini 4 Carbon |
|---|---|---|---|---|
| FrontierCode 2.0 | 49.1% (Xhigh) | 45.5% | 42.8% | 41.1% (Medium) |
| FrontierSWE v2 | 69.9% | 62.3% | 65.5% | 61.2% |
| AutomationBench | 51.4% | 42.5% | 41.4% | 53.9% |
| Terminal-Bench-Science 0.1 | 74.6% | 58.7% | 68.1% | 64.8% |
| OSWorld 2.1 (offline, partial) | 89.3% | 87.2% | 79.5% | 81.5% |
| AA-Briefcase v1.1 (rubric) | 66.1% | 61.2% | 52.0% | 61.8% |
| AA-Briefcase v1.1 (Elo) | 1785 | 1800 | 1569 | 1611 |
| Chartography (no tools) | 73.4% | 64.4% | 71.0% | 72.5% |
Three caveats belong next to that table. Anthropic says all scores use each model's highest available effort unless noted, but its FrontierCode row marks Carbon at Medium effort against Fable at Xhigh, so that gap is not like for like. The page does not say whether Anthropic ran the Carbon numbers itself or took them from Google. And Google's own post says Carbon ranks first on CursorBench 4.0, Perplexity's WANDR and AutomationBench, at more than 2 times lower cost per task than the next-best external model. Anthropic's table has a WANDR row with Fable 5.2 at 82.1 percent but no Carbon figure, so the two first-place claims cannot be checked against each other from what either company published.
The close rows matter more than the wide ones. On Chartography, Carbon trails Fable 5.2 by 0.9 points. On the AA-Briefcase rubric it trails by 4.3 and edges Opus 5.5 by 0.6. At a quarter of Fable's price through November 30, and half of it after, those are the rows where a buyer will ask what the extra money buys.
Our Take
Google is not trying to win the Fable tier on price. It is pulling the frontier down into the Opus tier. At $4 and $20, Carbon sits line for line on Opus 5.5, cache rate included, and Argon sits there too. That makes Google's pitch simple: the strongest model it sells costs what Anthropic charges for its daily driver. Anthropic's own launch page recommends starting with Opus 5.5 for most work, so this is the comparison that will decide real budgets, not Carbon against Fable.
The three-week window is the part I would not build a budget on. At $2 and $10 through November 30, Carbon is cheap enough that teams will run their evals on it now, at a price that ends three weeks after launch. Anyone migrating should price the move at $4 and $20 and count Carbon's longer reasoning in tokens, not per-token rates. Google's own guidance to use Argon for most work is a quiet admission that Carbon's per-task bill can run higher than its per-token price suggests.
For Anthropic, AutomationBench is the row to watch. It is the only row in its own table labeled business workflows, and the company published itself losing it to a model that costs half as much from December. That is an honest table. It is also a sign that the Fable premium now has to be justified on coding and science, where Anthropic's numbers show wide leads, rather than on general office automation.
Three Signposts for the Next 60 Days
First, whether Google holds the December 1 step to $4 and $20 or extends the introductory rate, a decision that will say something about how the three weeks went. Second, whether independent runs of AutomationBench, WANDR and CursorBench with all three vendors at matched effort settle the competing first-place claims. Third, whether Anthropic answers with another Fable price move or leans on Opus 5.5 at $4 and $20 as its head-to-head product against Carbon.
Sources: Google on Gemini 4 Carbon availability, Google introducing Gemini 4 Carbon, Gemini API pricing, Anthropic on Fable 5.2 and Mythos 5.2, Claude API pricing, OpenAI API pricing, The Next Web, Android Authority, Let's Data Science, SD Times.