Gemini 4 Carbon’s launch price has 20 days to run. Google made Carbon generally available on November 10 at $2 per million input tokens and $10 per million output, and the footnote on the announcement is the line to plan around: “After the introductory period expires on November 30, 2026, the price of $4 per 1M input tokens and $20 per 1M output tokens will apply.” The AI Studio pricing page shows the same switch on every tier.
That turns this week’s coverage around. VentureBeat’s headline on the release is that Carbon undercuts Claude Fable 5.2 and Opus 5.5, and through November that is true: a quarter of Fable 5.2’s $8 and $40, half of Opus 5.5. But any routing decision you make this month will still be running in January. On December 1, Carbon and Opus 5.5 have the same sticker price, so the useful comparison is the one at December rates.
At December prices, Carbon and Opus 5.5 cost the same per token
Here is the per-million-token comparison once the promotion ends, from Google’s pricing page and Anthropic’s model overview:
| Gemini 4 Carbon (from Dec 1) | Claude Opus 5.5 | Claude Fable 5.2 | |
|---|---|---|---|
| Input | $4.00 | $4.00 | $8.00 |
| Output | $20.00 | $20.00 | $40.00 |
| Cached input read | $0.20 | $0.20 | $0.20 |
| Batch input / output | $2.00 / $10.00 | 50% discount | 50% discount |
| Explicit cache control | No, implicit only | Yes | Yes |
The cache read row is the one most people skip. On a long agent loop where the shared prefix makes up most of the tokens, all three models pay $0.20 per million for that part of the bill. The gaps that remain are on uncached input and output, and that is where thinking tokens land: Google’s page counts output “including thinking tokens.”
Per token, then, the December choice between Carbon and Opus 5.5 is not a price decision. It comes down to which model finishes your tasks in fewer tokens and fewer attempts.
Per task, the cheapest sticker does not always win
Artificial Analysis runs every model through the same nine-evaluation suite and publishes the average cost per task. On today’s v5.0.1 pages:
- Claude Fable 5.2 at max effort: Intelligence Index 65, $26.54 per task
- Gemini 4 Carbon at high: 60, $7.76 per task
- Claude Opus 5.5 at max: 59, $10.04 per task
- GPT-6 Astra at max: 57, $6.24 per task
The Carbon page lists its price as $4 and $20, the post-November rate. If its per-task figure is computed at that listed price, $7.76 is already a December number, and Carbon beats Opus 5.5 on both score and cost at those settings.
The lower effort settings tell a different story. VentureBeat, reading the same leaderboard, reports that “at medium effort, Fable costs $3.80 and scores 56, while Carbon costs $6.59 and scores 58.” A model with twice the per-token rate comes in at about 58% of the per-task cost (our arithmetic), two index points behind. That is the effect we described in this morning’s Fable 5.2 guide: a model that reaches an answer in fewer reasoning tokens can be cheaper per task than one with a lower rate card. It cuts in Carbon’s favor at high effort and against it at medium, which is why you price your own workload at the effort level you will actually run.
Where each one wins on published benchmarks
No model sweeps the table. From VentureBeat’s comparison of Google’s, Anthropic’s and Artificial Analysis’s figures, plus Anthropic’s own launch table:
Carbon leads on Terminal-Bench 4.0 (65.2% at high, against Fable’s best listed 61.6%) and on Zapier’s AutomationBench (53.88% at high for an estimated $2.12 per task, against Fable 5.2 max at 51.45% and $4.64). Google’s post claims first place on CursorBench 4.0 and Perplexity’s WANDR “at more than 2X lower cost per task compared to the next best external model,” but VentureBeat notes the public tables it checked had no Carbon row for either, which leaves those two claims unverified for now.
Fable 5.2 leads on long, hard coding and science work: FrontierCode 2.0 (49.1% against Carbon’s 40.6%), FrontierSWE v2 (69.9% against 61.2%), and Terminal-Bench-Science (74.6% against 64.8%). Those are the long-horizon tasks Anthropic says justify its price.
Opus 5.5 leads on computer use: 87.2% on OSWorld 2.1 against Carbon’s 81.5%, per VentureBeat’s October table. Fable 5.2 is ahead of both at 89.3%, but at double the rate. VentureBeat’s November 10 report also notes Opus 5.5 still tops Artificial Analysis’s GDPval-AA v2.1 leaderboard for general knowledge work.
Read that as three lanes rather than a ranking. Terminal and workflow automation is where Carbon now has the best published evidence. Novel, multi-day engineering is still Fable’s case. Desktop automation and everyday document work have no published reason to leave Opus 5.5.
Latency is now Carbon’s argument, not its problem
When Carbon was announced, its slow start was the main caution in our pre-access guide. Against the Claude models it is now the faster one. Artificial Analysis lists time to first answer token, which includes reasoning, at 66.85 seconds for Carbon at high, 235.06 seconds for Fable 5.2 at max and 369.72 seconds for Opus 5.5 at max. Average time per index task is 25.6 minutes for Carbon, 25.1 for Opus 5.5 and 44.4 for Fable 5.2.
The settings are not matched (high against max), and a fair comparison would run each at the effort level you would ship. Our reasoning effort guide covers how to find that level. But if a person waits on the first visible token, a minute against four to six minutes matters more than an index point.
What the GA release still leaves open
Several things a procurement or platform team needs are not in the announcement:
- Where it runs. Google’s post names AI Studio, Android Studio and “Gemini for business.” It does not mention Vertex AI by name, and VentureBeat reports that Google “has yet to publish a country-by-country rollout schedule for Carbon.” Artificial Analysis lists one API provider for Carbon, against nine for Opus 5.5 and six for Fable 5.2. If your contracts run through Bedrock, Azure or a Vertex commitment, check before you plan on Carbon.
- Cache control. The pricing page says “Implicit caching only (explicit caching is not supported).” Anthropic lets you mark cache breakpoints yourself. With implicit caching, your hit rate depends on prompt structure and Google’s own matching, so measure it on real traffic before you assume the $0.20 rate on most of your input. Our caching guide covers the prompt layout that helps either way.
- Which Carbon you get. The pricing footnote “applies to the Carbon model intended for public access only.” The cyber variant stays in the Fairwind Program, and Google’s headline exploit scores were run without guardrails. The public model has stronger misuse safeguards: Google reports a per-attempt breach rate of 0.8% against Argon’s 3.1% in its own red-team testing. Security teams should not expect the public Carbon to match the cyber numbers from October.
- Honesty on your data. Google says Carbon passed 92% of a 375-task agentic honesty benchmark against Argon’s 70%, and declined to fill in missing data 98% of the time. Those are Google’s internal results. They answer the concern Artificial Analysis’s Omniscience numbers raised in October, that Carbon guessed more than Argon. Test it on your own data with declines scored separately from errors before you trust either figure.
For the next 20 days, Carbon at $2 and $10 is the cheapest way to run a frontier model on terminal and workflow automation, and it is worth putting into an eval against your own tasks this week. Budget it at $4 and $20. At that price it matches Opus 5.5 exactly, so Carbon gets your traffic only if it wins on finished tasks, wall-clock time or both. If the only reason is November’s discount, that reason is gone on December 1.