Skip to main content
Gemini 4 Carbon is Google’s frontier model for work that still tests the limit of Gemini 4 Argon: multi-step code, open-ended research, and dense documents with charts, tables, and diagrams. Public access opened on 10 November 2026. Google keeps Argon as the choice for most coding and knowledge work, and describes Carbon as the model that spends more tokens to explore, reason, and verify. The identifier on this page is gemini-4-carbon. The 1,048,576-token window is one budget, shared by input, reasoning, tool calls, and the reply. A single request stops at 65,536 tokens when streaming, or at 32,768 on a non-streaming HTTP call. Generation can continue up to the window, minus the input. Google’s guide also accepts Word, PowerPoint, Excel, and OpenDocument files. Agentic video understanding exists and is off by default.

​
When to use Carbon and when to use Argon

Google points to Argon for most software engineering and knowledge work. Carbon is for code that crosses packages, research without a script, and a dense document with a chart and a table.
Stay on Argon for everyday code, a short answer, and a document that fits in one read. Move to Carbon when the change crosses packages, the research has no script, or the data sits in a PDF with a chart and a table. For a short classification queue, LOW is enough.

​
What Google measured

The announcement does not publish the absolute score behind the three first-place claims. It publishes rank.
  • On CursorBench 4.0, Perplexity’s WANDR, and Zapier’s AutomationBench, Google says Carbon placed first.
  • On Multimodal Knowledge Agent, built with Artificial Analysis, it placed second.
  • On Hebbia’s end-to-end financial evaluation, it matched the best model tested. On Balyasny’s financial intelligence benchmark, it ranked at the top of 23 configurations.
  • On an internal set of 375 agentic honesty tasks, it passed 92%, versus 70% for Argon. It refused to fill in missing experimental data 98% of the time and flagged a methodological flaw 68% of the time.
Anthropic published 53.9% for Gemini 4 Carbon on the AutomationBench column of the Claude Fable 5.2 table. Google says Carbon leads Zapier’s AutomationBench and does not publish that score. Do not treat the two as the same test.

​
What Artificial Analysis measured

On the high variant, Artificial Analysis records an Intelligence Index of 60, 4th in the sample they published. The same run shows 66.85 seconds to the first answer token, 25.6 minutes per index task, and 80 output tokens per second. That run produced 205 million output tokens. The page dates this variant to 28 October 2026, before public access on 10 November.
The 60 and the time to the first answer are for the high variant, not MEDIUM thinking, which is the default on gemini-4-carbon.

​
Where it does not lead

  • Google still points to Argon for most code and knowledge work. Carbon spends more tokens in the same family.
  • On Multimodal Knowledge Agent, the announcement places Carbon second, not first.
  • On the high variant Artificial Analysis measured, the first answer is slow and an index task takes tens of minutes.
  • temperature, top_p, and top_k are ignored. Effort is controlled with thinking_level. MINIMAL and thinking_budget return an error.
  • A single request still stops at 65,536 tokens when streaming. Going past that requires continuation.

​
Where it creates the most value

​
Code that Argon still does not close

The change crosses packages, the cause sits in another subsystem, or the architecture does not fit in one pass. Carbon is the next step, with more verification and more tokens. Expected result: the diff inside the scope, the tests for that step, and what was left out.

​
A dossier that becomes a spreadsheet, report, or deck

The material mixes PDFs, tables, charts, and prose. Google describes Carbon as following a schema and delivering the spreadsheet, report, or presentation with the figure traceable. Expected result: the file you asked for, the source of each number, and a gap when the document does not have the figure.

​
Research without a fixed script

The question is scientific, financial, or methodological, and the answer has to test the hypothesis instead of filling in what is missing. On Google’s internal set, Carbon refused to invent experimental data more often than Argon. Expected result: what was measured, what is a hypothesis, and what the material does not say.

​
How to apply it in Tess

1

Move to Carbon only when Argon falls short

Use Argon for everyday code and a document that fits in one read. Choose Carbon for a change that crosses packages, a PDF with a chart and a table, or research without a script.
2

Define what done means

State the deliverable, what must not change, and the check that ends the task. Carbon tends to investigate more. Ask for the final file, not a plan of what it would do next.
3

Define tools and approvals

Enable only the tools the task needs. Require confirmation before sending, publishing, or changing data.
4

Choose the thinking level

LOW to classify and extract. MEDIUM, the default, for general code, tools, and documents. HIGH for many files, a proof, or an architecture.
5

Sample before you scale

Run one full case, check the result, and only then repeat the flow.

​
Three playbooks to start

​
1. A change Argon did not close

Input: the repository, what already failed on Argon, and the tests for this step.
Done when: the tests for this step pass, and what was left out is written down.

​
2. A dossier with a final file

Input: the PDFs or spreadsheets, the deliverable schema, and the question.
Done when: the file exists, and each number points to a source or to a gap.

​
3. Research with an explicit gap

Input: the question, the data, and what counts as evidence.
Done when: the answer separates what is in the material from what was inferred.

​
Limitations and best practices

  • Default thinking on gemini-4-carbon is MEDIUM.
  • The announcement says Carbon uses more tokens than Argon to verify the work.
  • A 1,048,576-token context does not replace selection. Output can use only what remains after the input.
  • Put the stable prefix at the start of the prompt. Implicit cache looks at prefixes from 4,096 tokens.
  • Google reports a 0% prompt-injection success rate on Gray Swan’s IPI benchmark and, in the same announcement, says red teaming still finds a more sophisticated attack. Review a sample anyway.
  • Keep a human approval for financial, legal, clinical, or high-impact decisions.
See also: Gemini 4 Argon · Claude Fable 5.2 · Models and Costs · Gemini 4 Carbon is now available · Developer guide · Artificial Analysis · LLM Stats.