Gemini 4 Carbon pricing and availability — a quarter of Fable 5.2 per token, 15% apart per task

Google's top model lost its defenders-only label after two weeks. On Tuesday, November 10, 2026 (US time), Google said Gemini 4 Carbon is available to paying Gemini API developers and enterprise customers and is rolling out to Google AI Ultra subscribers. When Carbon was first announced on October 28, only a subset of cybersecurity defenders and partners in Google's Fairwind Program could use it, and no date was given for anyone else.

The price is what Google previewed then: $2 per million input tokens and $10 per million output tokens. What is new is an end date. Google's price list says those rates run through November 30, 2026 and become $4 and $20 on December 1. A day earlier, on November 9, Anthropic released Claude Fable 5.2 at $8 and $40. Per token, Carbon costs a quarter as much. Measured as the cost of finishing a task, the gap is much smaller. The numbers are set side by side below.

1. What changed between October 28 and November 10

The announcement was signed by Koray Kavukcuoglu, senior vice president at Google DeepMind. Google calls Carbon its strongest model, yet says Gemini 4 Argon remains the right choice for most software engineering and knowledge work. Carbon is for harder tasks and spends more tokens on reasoning and verification.

Who is left out is equally clear. The price list states that Carbon is not available on the free tier. Argon went to both Google AI Pro and Ultra subscribers on October 13; the Carbon post names Ultra only. According to VentureBeat, Ultra costs $99.99 or $199.99 a month in the United States. Google has not published a country-by-country schedule, and even Ultra subscribers are getting access in stages.

Two things shipped alongside. One is a preview of /swarm in Antigravity, Google's coding tool: a coordinating agent splits a project among worker agents, and separate verifier agents run tests and review the changes. Each role can use a different model, so Carbon can handle design and auditing while lighter models do routine implementation. The Antigravity team warns that parallel multi-agent work burns through usage allowances far faster than ordinary chat. The other is a research program: selected scientists get Argon and Carbon free for six months on Antigravity's highest usage tier.

2. Pricing — $2 until November 30, $4 after

These are the rates on Google's price list. The model ID is gemini-4-carbon.

Output billing includes the thinking tokens the model spends before answering. Carbon is built to think more than Argon, so identical per-token rates do not produce identical bills for the same job.

Here is how the list prices compare with other models released in the same stretch. The Fable 5.2 and Opus 5.5 figures match Anthropic's announcement and VentureBeat's comparison table.

For a small service using 10 million input tokens and 2 million output tokens a month, the difference looks like this.

The discount window lasts three weeks from the announcement. After December 1, Carbon costs the same per token as Opus 5.5, and its advantage over Fable 5.2 narrows from a quarter to a half. That arithmetic assumes two models use the same number of tokens for the same work. They do not.

3. Performance — three first-place claims and what outside data shows

Google says Carbon ranks first on CursorBench 4.0, Perplexity's WANDR and Zapier's AutomationBench, at less than half the cost per task of the next best external model. The three claims stand on different footing.

Only AutomationBench can be checked against numbers. Zapier's report shows Carbon passing 354 of 657 tasks (53.9%) on its high setting at an estimated $2.12 per task, and 53.0% at $1.65 on medium. In the same report Argon scores 51.3%, Claude Sonnet 5.5 44.7%, Opus 5.5 42.5% and GPT-6 Astra 41.4%. Anthropic's own launch table of November 9 also shows Carbon (53.9%) ahead of Fable 5.2 (51.4%) on this line. It is a rare case where a vendor's claim and a rival's table point the same way.

The caveats are in Zapier's report too. The run took place before release, on October 26 and 27, once per effort setting, with no confidence interval. Costs are list-price estimates, not billed amounts. Tasks run inside 47 simulated apps. And Fable 5.2 is absent, because the report predates it. VentureBeat puts Fable 5.2 at max effort at 51.45% and $4.64 per task.

For a broad measure Google does not control, there is the Artificial Analysis Intelligence Index. This is the leaderboard as of November 12.

On October 28 Carbon was first on this index. Two weeks later it is fourth, behind three Fable 5.2 configurations. Its score moved from 59 to 60 because, according to VentureBeat, one scoring component was updated and models were regraded.

This table is where the price story turns. Fable 5.2 costs four times as much per token, yet at the same high setting the cost per task is $9.08 against $7.76. Carbon is about 15% cheaper and Fable scores two points higher. Fable 5.2 on medium costs $3.80, less than Carbon on medium at $6.59, while scoring two points lower. At the other end, Fable 5.2 on max costs $26.54, or 3.4 times Carbon. Carbon uses more tokens per task, which absorbs much of its per-token advantage. These costs reflect the introductory rate; once the rate doubles on December 1, the same comparison may tip the other way. One more line stands out: Argon, the model one step down, costs more per task ($9.95) than Carbon ($7.76).

Coding results are split. In Artificial Analysis head-to-head figures compiled by VentureBeat, Carbon on high leads Terminal-Bench 4.0 at 65.2% against 61.6% for the best listed Fable 5.2 run, while Fable leads FrontierCode 2.0 at 49.1% against 40.6%. Google's own model card lists Carbon at 61.2% on FrontierSWE 2, behind GPT-6 Astra (65.5%) and Opus 5.5 (62.3%), the same Carbon figure that appears in Anthropic's table. On the independent Vals AI index, as reported in our November 11 article, Carbon was third at 70.45%, after Fable 5.2 (71.31%) and Sonnet 5.5 (70.85%).

In short, Carbon leads on workflow automation and terminal tasks, and Fable 5.2 leads on long-horizon coding and the composite index. Google's claim of less than half the cost per task holds when top settings are compared and does not hold at matched settings.

4. How much of the cyber capability is actually open

The security scores Google highlighted in October (88.2% on SEC-bench Pro, 54.8% on ExploitGym, 71.0% on CWE-bench) were measured with guardrails off. That is not the model that opened on November 10. The public version carries cyber safeguards, and the version with fewer restrictions stays inside the Fairwind Program under the name Carbon for cyber. Defenders still apply and are vetted. Giving a model to defenders first and then releasing a safeguarded version to everyone is the pattern known as trusted access, and it mirrors Anthropic's split between Fable (public) and Mythos (vetted organizations).

Google offered these safeguard figures in support of the wider release. All are Google's own or from tests Google commissioned.

A less prominent number deserves equal attention. As relayed by Unite.AI from Google's safety report, when 25 high-severity objectives were each attacked 200 times, 40% of the objectives were eventually breached. A single attempt succeeds less than 1% of the time; persistent retries got through on four objectives in ten. The same report says that without external safeguards Carbon cannot be ruled out from reaching Cyber Uplift Level 1 on Google's scale, and judges that strengthened safeguards bring residual risk to an acceptable level. Google also wrote that robustness to indirect prompt injection remains an area for continued improvement, despite the 0% result.

On the defensive side, Google says security firm Wiz used Carbon to find previously unknown vulnerabilities at two major software-as-a-service providers in under two hours. The providers are not named.

5. A day later: the Decisions API, a model that writes nothing

On November 11 (US time) Google separately launched the Gemini Decisions API. It sits at the opposite end of the range from Carbon. The model, Gemini Decision 1 (gemini-decision-1), does not generate text token by token. It scores a set of predefined options in one pass and returns probabilities.

When a large model like Carbon works as an agent, it makes many small calls along the way: which tool to use, whether a document is relevant, which model should take a request. Calling a frontier model for each is slow and expensive. The Decisions API is the component meant for that slot. Google says it is nearly five times cheaper than Gemini 3.5 Flash-Lite across its own benchmark suite and cut latency for everyday queries in Gemini Notebook by up to 40%. Those are Google's measurements; no independent tests exist yet. The latency figure is not a promise for a full 1,000-question request or a 256K-token input.

6. What remains open

The most practical variable is the calendar. The window for testing Carbon at $2 and $10 closes on November 30. After that it is priced like Opus 5.5 per token, and the choice turns on how many tokens each model needs to finish the same work. That has to be measured on your own workload, because index tasks are not your tasks.

The security scores and Fairwind vetting terms from the limited release are in our October 30 piece, 「Gemini 4 Carbon launch」. Competing prices and independent measurements are in the November 11 piece, 「Claude Fable 5.2 released」. The logic of giving strong models to defenders first is explained in 「What is trusted access in AI?」, and the model one step down, whose price rises on the same day, is covered in 「Gemini 4 Argon launched」.