Google has opened Gemini 4 Carbon to every paying developer, ending a two-week window in which the company’s most capable model was limited to vetted cybersecurity defenders. The catch is written into the price list: the introductory rate runs only through 30 November. On 1 December, the cost of input tokens doubles from $2 to $4 per million, and output tokens double from $10 to $20 per million.

The change matters beyond a line on a cloud bill. Carbon is the model Google is positioning for long, multi-step work, debugging a large codebase, auditing software for security flaws, or researching an open-ended question, where a model may spend minutes thinking before it answers. Those workloads are exactly the ones that burn through tokens, and token prices are how AI labs charge for them.

What Google actually announced

Carbon is now available on the paid tier of the Gemini API, in Google AI Studio and in Android Studio. Businesses can reach it through Gemini for business, Google Antigravity and CodeMender. Google AI Ultra subscribers can use it in the Gemini app, AI Mode in Search and Antigravity. The model ID is gemini-4-carbon.

It is not available on the free tier. That is a deliberate change from the usual Gemini playbook, where a free tier often serves as the funnel. Carbon is aimed at people who are already paying for inference and who care more about whether the model can finish a job than whether the first million tokens are free.

Google announced Carbon in late October and gave it to cybersecurity defenders through its Fairwind Program. The wider release places it alongside Gemini 4 Argon, the faster model Google says is the right default for most software engineering and knowledge work because it spends fewer tokens per turn. Carbon is the specialist: slower, more deliberate and more expensive to run.

What makes Carbon different from Argon

Both models share a unified one-million-token context window, built-in tools and the same per-token price during the introductory period. The difference is how much thinking they do. Carbon spends more tokens per turn exploring hypotheses, checking its work and testing edge cases. Argon is tuned for faster responses and lower token spend on routine coding and everyday knowledge work.

That trade-off is the whole product. A model that thinks longer can solve problems a faster model gets wrong, but it also produces more billable output. Google exposes the dial through a thinking level setting: low, medium or high. Medium is the default. Google’s own guidance is blunt: run your evaluations at all three levels and pick the lowest one that meets your quality bar.

Carbon also supports long decode, which lets a single response run past the usual 64,000-token limit, up to the one-million-token context window minus the input. For outputs above 64K, developers have to enable background execution. That is useful for long code generations or detailed reports, and dangerous for costs if left on by default.

The price change, in plain numbers

Until 30 November, Carbon costs $2 per million input tokens and $10 per million output tokens. Cached input, material the model has already processed and can reuse, is discounted by 95% to $0.10 per million tokens. From 1 December, the standard rates are $4 and $20, with cached input at $0.20 per million.

A worked example makes the deadline concrete. Suppose an agent run sends 100,000 tokens of repository context and produces 20,000 tokens of output. At the introductory rate, the input costs $0.20 and the output costs $0.20, for a total of $0.40. At the December rate, the same run costs $0.40 for input and $0.40 for output, or $0.80.

Double a single run and you double a fleet. At 5,000 runs a day, the bill moves from roughly $2,000 a day to $4,000 a day. Over a 30-day month, that is about $60,000 versus $120,000. The exact figure depends on how much of the input is cached and how much thinking the model does, but the direction is unavoidable.

Cached input changes the shape of the bill. A team that reuses a large system prompt or a stable codebase context can pay a fraction of the headline input rate. The workloads that get expensive are the ones that generate a lot of output: code generation, long analyses and multi-file refactors. Output is priced at five times input, and on Carbon, thinking tokens count as output.

The part the headline misses

The sticker price doubles. The bill for a like-for-like workload can rise by more than that.

Carbon is designed to spend more thinking tokens per turn than Argon. Google bills every token the model processes, including thinking tokens. Swap Argon for Carbon without changing prompts, output limits or thinking levels, and the same request may generate more billable output even before the December rate card applies. The effective increase is the price change multiplied by the change in token use.

That does not make Carbon a bad deal. Google 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. It also ranked second on a new Multimodal Knowledge Agent benchmark built with Artificial Analysis, at more than 6× lower average cost per task than the top model. If Carbon finishes in one pass what a cheaper model needs three attempts to complete, the cost per finished task can fall even though the cost per token rises.

But those are Google’s benchmarks, run on Google’s chosen tasks. They are a reason to test Carbon, not a reason to assume your bill will behave the same way. The only number that matters is your cost per completed job, measured on your own workload.

Fine print that affects the bill

Several details are easy to miss and expensive to discover later.

Carbon is not available on the free tier. Grounding with Google Search is still billed per query on top of token costs. Explicit caching is not supported on Carbon, so teams that create and reference cached content need to remove those calls; implicit caching is on by default for prompts of at least 4,096 tokens and delivers the 95% discount on cache hits. Sampling parameters such as temperature, top_p and top_k are ignored. And max_output_tokens caps thinking plus output combined, not just the visible answer, so a limit that was generous on an older model can cut Carbon off mid-reasoning.

There is also a latency cost. Google warns that hard prompts can take several minutes and recommends HTTP timeouts of at least 15 minutes, or background execution for long outputs. A model that is cheaper per successful task can still be more expensive operationally if it forces changes to timeouts, retries and queue design.

How to decide before 1 December

The deadline is a forcing function, not a reason to rush a migration. A sensible sequence looks like this.

First, benchmark Carbon against Argon and against whatever model you use today on a representative set of tasks, not a benchmark suite. Measure success rate, retries, latency and total tokens per completed task. Second, test all three thinking levels. High is not automatically better; it is automatically more expensive. Third, route work by difficulty. Use Argon for routine feature work and interactive coding, and reserve Carbon for root-cause debugging, architectural design, security review and dense multimodal analysis.

Fourth, restructure prompts so the stable part clears the 4,096-token threshold for implicit caching. Fifth, set output limits deliberately and use background mode only where you need it. Sixth, put a budget alert on the project before the price change, not after.

For teams that already know Carbon is central to their product, the next three weeks are a chance to lock in usage patterns at the lower rate. For teams that are curious, the introductory price is a cheap way to find out whether the model’s extra reasoning is worth paying for. Either way, the decision should be based on cost per task, not cost per token.

What to watch

Three things. Independent evaluations of Carbon on real repositories, not vendor-selected benchmarks. Whether Google extends the introductory price if demand softens, which would tell you how much of the deadline is marketing. And how OpenAI and Anthropic respond; a price cut from a rival would change the arithmetic faster than any benchmark.

There is a broader signal too. Google is now charging a premium for reasoning depth while keeping the cheaper model close enough that switching is a one-line change. That is a bet that developers will pay for reliability when a task is hard and route around it when a task is easy. The deadline tests how many teams agree.

The bottom line

Gemini 4 Carbon is available to anyone with a paid Gemini API key, and it is the most capable model Google has shipped. The introductory price of $2 per million input tokens and $10 per million output tokens ends on 30 November. From 1 December, the same tokens cost $4 and $20.

If you are already using Carbon, audit your thinking levels and output limits now. If you are not, the next three weeks are the cheapest window you will get to find out whether it earns a place in your stack. The discount is real, but it is also a countdown.

Have a tip? Email [email protected] or use the contact form and choose Editorial. We do not promise anonymity.

Sources: Google; Google AI Studio documentation.