Gemini 4 Carbon Explained: Access, Benchmarks and When to Wait

Gemini 4 Carbon is not publicly available as of 29 October 2026, so keep shipping on your current model. Google DeepMind announced it on 28 October for demanding engineering and cybersecurity work; a leaderboard lead is a reason to prepare an evaluation, not delay a release. Google's announcement sets out a restricted rollout.

What Gemini 4 Carbon Is Built to Do

Carbon targets work that takes repeated attempts to finish: sustained coding, difficult engineering and cyber defense. Google says its internal teams have seen it work for days on debugging and patching complex codebases.

Long-horizon work means keeping a task on track across many steps. Think of a maintainer following a bug from a failing test through several files, trying a fix, discovering a second failure and revising the patch. An agent is the software around the model that lets it take actions, such as reading files and running tests. Success depends on the whole arrangement, including the tools and the checks that decide whether the job is done.

For a founder, the useful question is whether this can finish a recurring maintenance job that currently consumes an engineer's attention. A polished answer to a coding question is a different deliverable from a patch your team accepts.

Who Can Use Carbon Now?

Google's sequence is selected cyber defenders first during announcement week, additional trusted partners through Fairwind next, and wider product access after feedback and guardrail improvements. It promises wider access “as soon as possible,” without a date.

The Fairwind Program is Google's vetted route for defensive organizations. It prioritizes governments, critical infrastructure and core technology platforms. Its application form is an eligibility review, not an instant public-model signup. A security founder can apply if the organization fits; an ordinary application builder should not make delivery depend on acceptance.

The program's public technology descriptions still name Argon. Use the Carbon announcement for the rollout sequence and Fairwind's page for the program rules. Neither an application nor existing program membership proves that Carbon has been enabled for your organization.

The deployment gap: Google's announcement supplies no public Carbon model ID, context limit, app tier or general-release date. A benchmark's input length does not establish an API limit. Do not fill those fields by copying Argon's documentation.

These sources and Google's evaluation methodology were checked on 29 October 2026. This is a dated access assessment; no hands-on Carbon testing was performed for this article.

How Carbon Relates to Argon

Gemini 4 Argon remains the available Gemini baseline. Google made Argon available on Google Cloud on 13 October, and its Gemini API pricing documentation lists paid access. Carbon builds on Argon, but its announcement does not replace the model your application can call today.

That distinction matters for a solo builder maintaining an existing integration. Continue improving your prompts, tool permissions and acceptance checks against Argon. Save the difficult tasks it fails so a future Carbon evaluation has something meaningful to beat.

For a CTO, separate the migration decision from the release announcement. Reusing a test set is straightforward; assuming identical API behavior is not. Confirm the eventual model's documented inputs, outputs, tools and operational limits before treating it as a drop-in replacement.

The Gemini 4 Argon explainer covers its current access routes and operating limits. Those belong to Argon until Google documents otherwise for Carbon.

The Benchmarks Google Publishes

Google's published comparison shows Carbon ahead of Argon across the listed evaluations, but it does not prove Carbon wins every task or every competitor comparison. The table below transcribes Google's benchmark sheet. These are published results, not measurements made for this article.

The cybersecurity rows describe different stages of defense: find a flaw, establish whether it is exploitable, and fix it. Do not average the percentages into a security score. A patching benchmark is the closer starting point for a team evaluating maintenance work; a vulnerability-finding result answers a different question.

The Test Conditions Matter

Google's evaluation methodology generally uses the highest thinking settings. It sources AA Index and Terminal-Bench results from Artificial Analysis and CWE-bench from Collinear AI, while computing other results itself. The sheet mixes evaluation sources.

For example, Google's SEC-bench Pro result uses a 183-task JavaScript-engine split. Its ExploitGym evaluation uses the Antigravity harness, the software that runs the agent's tools, without internet and with mitigations off. OSWorld reports an offline subset and the maximum across runs. These conditions matter when deciding what a score can support.

A product running with different tools, permissions or limits needs its own evaluation. Published benchmark results do not promise that the restricted capabilities used in a test will be exposed through the product your business eventually buys.

What Is Overhyped About “#1”

“#1” identifies a particular ranking, not a universal winner. Artificial Analysis lists Carbon High at the top of Intelligence Index v5.0 with 59, while explicitly marking the absence of a publicly available API. Strong evaluation access and public customer access are different facts.

Google's own sheet also shows exceptions: Carbon's 81.5% on the OSWorld offline subset trails Claude Opus 5.5, Anthropic's model, at 87.2%. That is enough to reject the claim that Carbon leads every workload. Choose the evaluation closest to your work before choosing a model.

Do not turn the table into a Carbon-versus-OpenAI GPT-6.1 Terra verdict either. Its OpenAI column is GPT-6 Astra, a different model. A result for Astra cannot answer whether you should replace Terra in your coding workflow.

Finally, Google's reported 0.0% prompt-injection attack success applies to the evaluated attacks. Prompt injection means instructions hidden in material an agent reads that try to redirect it. Keep permission boundaries, test review and approval for consequential actions even when a model posts a strong result.

What Builders, Operators and Buyers Should Do

Builders should keep the model that is delivering acceptable work. If your Gemini application runs on Argon, stay there while preparing the comparison. If your workflow already uses Opus 5.5 or GPT-6.1 Terra successfully, Carbon's announcement alone is no reason to stop or migrate. The current model guide covers the broader choice.

Operators should make long tasks reviewable. A mid-market CTO can nominate a recurring regression fix with a known failing test and a clear owner. A senior operator can require a task log showing what changed, which checks ran and what remains unresolved. More time spent working only helps when the result is correct and the reviewer can establish that efficiently.

Buyers should budget beyond the introductory period. Tokens are the units billed for model input and output; a longer run can consume more of them even if its final patch is small.

Gemini 4 Carbon price: $2 per million input tokens and $10 per million output tokens initially; cached input is 95% off input token price; after introduction, $4 per 1M input tokens and $20 per 1M output tokens. Google's post and footnote.

Carbon's post gives no introductory expiry date. Argon's documented Standard-rate change starts 1 December 2026; that is an Argon budgeting deadline, not a Carbon release or pricing deadline. Keep the detailed plan comparison on the Gemini pricing page.

Your Next Move This Week

Assign an owner to a reusable test set now. For a funded founder, the deliverable is a defensible adoption decision once access arrives. For a solo builder, it is a saved set of difficult jobs that makes the next model release quicker to assess.

1. 1

   Save representative work

   Collect a regression patch, a multi-file change, an authorized security fix and an ambiguous request that should trigger clarification. Include ordinary work as well as difficult failures. Remove credentials and unnecessary customer data from the saved fixtures.

2. 2

   Define acceptance before running

   Record the starting revision, permitted files, available tools, tests that must pass and conditions requiring a stop. For the regression, require both a passing reproduction test and no unrelated changes. A plausible explanation does not count as a finished fix.

3. 3

   Record your Argon baseline

   Run the set with the model you currently use. Save the patch, test results, billed usage, retries, interruptions and reviewer effort. These are proposed checks, not results measured for this article. Budget recurring Gemini work at the rate that will apply after introduction.

4. 4

   Make Carbon earn the switch

   When your account has access and Google documents the interface, repeat the same tasks with comparable tools and permissions. Expand use only if accepted results improve without an unacceptable increase in review, latency or cost. Otherwise, keep the working baseline.

For practical model-adoption decisions as access changes, subscribe to the newsletter.