Based on each vendor’s official docs, checked November 12, 2026

Gemini 4 Carbon in coding tools: API pricing, Cursor and CursorBench

Gemini 4 Carbon is Google's most capable Gemini reasoning model, intended for complex software engineering, long agentic work, and difficult science or math problems. This guide covers how to use it in coding tools, with figures checked on November 12, 2026 against Google's Gemini API pricing and model docs, Google's November 10, 2026 availability post, Cursor's Gemini 4 Carbon model page, Cursor's CursorBench 4.0 leaderboard, and Google Cloud's Agent Platform pricing page. It does not invent support in tools Google or Cursor have not documented, and it does not treat a benchmark score as a forecast for your repository. Google still describes Gemini 4 Argon as the better default for most software engineering and knowledge work; Carbon is the model it points at when a task needs deeper reasoning and, in Google's words, uses more tokens than Argon to explore thoroughly.

1. 1. Confirm where Carbon is actually available

   Google's November 10, 2026 availability post and the Gemini API model docs are the source for access, not a third-party recap. On the Gemini Developer API, the public model id is gemini-4-carbon, and it is not available on the Free tier: every token the model processes is billed, including thinking tokens and, when you use grounding, internal tool instructions and function declarations. Paying developers can use it in Google AI Studio and Android Studio. Google AI Ultra subscribers can try it in Google Antigravity, the Gemini app, and AI Mode in Search. Enterprises can use it in Gemini for business. A separate scientists' program can include Carbon in Antigravity's highest usage tier for six months, subject to eligibility, capacity and selection. Cybersecurity defenders apply separately to the Fairwind Program for Carbon for cyber. In Cursor, Carbon is a third-party model on the model list, with three effort levels: high, medium (the default) and low. This guide does not claim Claude Code, Codex or GitHub Copilot expose Carbon unless those vendors document it.

2. 2. Choose Carbon or Argon before you worry about the list price

   On the Gemini API paid Standard tier, Carbon and Argon currently share the same published introductory rates: $2 per million input tokens and $10 per million output tokens through November 30, 2026, then $4 and $20 starting December 1, 2026. Output pricing includes thinking tokens. Context caching on that tier is $0.10 per million tokens through November 30 and $0.20 from December 1, and Google documents implicit caching only; explicit caching is not supported. Because the list prices match, the practical difference is how many tokens each model uses. Google's availability post says Argon remains the better choice for most software engineering, while Carbon is built for harder, multi-step coding and open-ended research and uses more tokens than Argon to run extended verification loops. Start with Argon on everyday work, and switch a single hard task to Carbon when you have a reason to spend the extra tokens.

3. 3. Set thinking or effort explicitly

   The Gemini API documents thinking support at low, medium and high for gemini-4-carbon. Cursor documents the same three effort levels and says medium is the default; raise it to high for the hardest tasks, or drop to low for faster responses on simpler ones. Cursor also says Carbon has access to all agent tools when used in Cursor. Do not assume a default in one product matches the other. After you change the level, run one representative task and check the token usage in that product's usage view before you make Carbon the default for a whole week of work.

4. 4. Read the rate that will apply after November 30, 2026

   Write down both the current introductory rate and the December 1, 2026 rate before you commit a workload. On the Gemini API Standard tier that is $2 input / $10 output through November 30, then $4 / $20. The Batch and Flex tabs are $1 / $5 through November 30 and $2 / $10 from December 1, with cache at $0.05 then $0.10. Priority is $3.60 / $18 through November 30 and $7.20 / $36 from December 1, with cache at $0.18 then $0.36. Google's availability post adds an important limit: after the introductory period expires on November 30, 2026, the $4 / $20 price applies to the Carbon model intended for public access only. Cursor bills Google's introductory rates through November 30, 2026 at $2 input, $0.10 cache read and $10 output per million tokens. From December 1, 2026, Cursor documents $4 input, $0.20 cache read and $20 output. Carbon is not part of the usage included with Cursor plans: on The X Subscription it is billed from prepaid credit, and on legacy Cursor plans it draws from the Other Models pool.

5. 5. Read CursorBench as Cursor's own leaderboard, not a guarantee

   CursorBench 4.0, checked on November 12, 2026 on Cursor's official leaderboard, evaluates agents on ambiguous, multi-file tasks from real Cursor sessions. Higher scores are better. Gemini 4 Carbon High is first at 60.4% ($9.22 cost per task, 39,960 tokens, 71 steps). Carbon Medium is second at 60.1% ($8.49, 37,067 tokens, 67 steps). Fable 5.2 Max is third at 60.0% ($32.79). Carbon Low is fourth at 58.6% ($11.89). Like-for-like, Fable 5.2 High is 56.9% ($9.70) and Fable 5.1 Max / High remain 51.8% and 49.2%. Cursor notes that small score differences may not be statistically meaningful, and that average cost per task applies each model's published per-million-token pricing to the tokens used. These are Cursor's own results, not an independent audit, and they are not a forecast of your repo's outcome. Google's availability post also says Carbon ranks first on CursorBench 4.0; this guide reports both that claim and the board itself.

Gemini API limits and what is billed

The official model page for gemini-4-carbon, checked November 12, 2026, lists a unified context window of 1,048,576 tokens and a default synchronous max output of 65,536 tokens, or up to 1,048,576 tokens with long decode. Supported inputs are text, image, video, audio and PDF; output is text. Documented capabilities include code execution, computer use, function calling, structured outputs, file search, URL context, grounding with Google Search and Maps, and implicit context caching. Image generation and the Live API are not supported. Batch, Flex and Priority consumption options are supported. The stable model id is gemini-4-carbon.

Every token counts on the paid API: thinking tokens, and when you use grounding, internal tool system instructions and function declarations. Grounding with Google Search on Gemini 3 and newer models includes 5,000 free search requests per month, shared across those models, then $14 per 1,000 requests, subject to a documented daily cap of 1,500 requests per day. A customer request may result in more than one Search query, and each query is charged. Do not assume the Free tier in AI Studio will serve Carbon; the pricing page states it is not available there.

Google Cloud published rates, with the promotional footnote

Google Cloud's Agent Platform pricing page, checked November 12, 2026, lists Gemini 4 Carbon through November 30, 2026 at $2 per million input tokens and $10 per million output tokens in the global region, with cached input at $0.10. Non-global regions are $2.20 input, $0.11 cached input and $11 output. Starting December 1, 2026, the same table lists $4.00 input and $20.00 output globally ($0.20 cached input), and $4.40 / $22.00 non-global ($0.22 cached input). A Priority table lists $3.60 / $18 through November 30 and $7.20 / $36 from December 1 in the global region.

Those Carbon rows carry an asterisk. The official footnote says promotional pricing is provided through 50% credits back on net spend on select models within a given period, and that promotional pricing does not apply to Gemini 4 Carbon Cyber Permissive or Gemini 4 Argon Cyber Permissive. This guide reports that footnote rather than trying to recompute a net unit price for your project. Google Cloud's model page also notes that Carbon is available to verified enterprise customers in select regions. A separate platform note says that when a Gemini 4 completion exceeds 32K tokens, retrievable token usage can be charged even if the request later fails due to a client or server timeout, effective November 13, 2026.

Antigravity, /swarm and consumer access

Google is previewing /swarm in Google Antigravity, which it describes as turning a single prompt into a coordinated team of agents. The availability post says developers can assign Carbon to roles that need the deepest reasoning, such as architectural design, coordination and auditing, and pair it with faster models in parallel workspaces. This guide does not invent a /swarm syntax beyond that official description, and it does not treat the demo projects in the post as templates you must reproduce.

If you subscribe to Google AI Ultra, official access includes Antigravity, the Gemini app and AI Mode in Search. That is a plan entitlement, not a Cursor entitlement, and it is not a Free-tier API grant. Scientists and researchers can apply for a rolling program that includes Carbon and Argon in Antigravity. Cyber capabilities that Google highlights for Fairwind, including the security-benchmark figures in the launch post, belong to that program; they are not automatically the public coding-tool experience.

Tips

• Use Argon for everyday coding unless you have a reason to spend Carbon's extra tokens.
• Write down both the through-November-30 rate and the December 1 rate before you attach Carbon to a production job.
• In Cursor, start from medium effort and raise a single hard task to high rather than leaving Max-equivalent effort on all day.
• Treat CursorBench gaps of a few tenths of a percent as Cursor's own scores, not a reason to switch your whole team.
• On Google Cloud, read the 50% credits-back footnote and the Cyber Permissive exclusions instead of assuming the table is a simple list price.

Next steps

Sources

• https://aistudio.google.com/docs/pricing · checked 2026-11-12
• https://aistudio.google.com/docs/models/gemini-4-carbon · checked 2026-11-12
• https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-carbon-now-available/ · checked 2026-11-12
• https://cursor.com/docs/models/gemini-4-carbon · checked 2026-11-12
• https://cursor.com/cursorbench · checked 2026-11-12
• https://cloud.google.com/model-garden/pricing · checked 2026-11-12
• https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/gemini/4-carbon · checked 2026-11-12