# Gemini 4 Carbon\: What Google Announced and Who Can Use It

> What Google says about Gemini 4 Carbon\'s capabilities\, safety evaluations\, availability and planned API pricing\, and what remains unknown\.

Google has announced Gemini 4 Carbon, a new frontier model aimed at demanding software engineering and cybersecurity work. Its first users will be a subset of cyber defenders, with access expanding gradually to trusted partners. Developers, businesses and consumers will have to wait: Google has not given a date for their rollout. [Google's announcement](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/) was published on October 28, 2026.

The distinction matters. Carbon is an announced model, but it is not yet a generally available replacement for Gemini 4 Argon. The figures below are Google's reported results, not measurements we have independently reproduced.

## What Carbon is designed to do

[Google describes Carbon](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/) as its strongest model, built for sustained, multi-step work. The company says its internal teams have seen it work for days on engineering problems, including tracing and patching subtle bugs across large codebases and low-level systems. That is a claim about observed internal workflows, not a promise that every task will run autonomously for days or succeed without supervision.

The announcement also points to full-stack web development, multimodal understanding and scientific tasks. Google reports a score of 1,769 and fourth place on WebDev Arena, an 89-point improvement, according to [Google’s evaluation](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/). Leaderboard positions and scores can change, and an arena ranking does not by itself establish how well a model will perform on a particular production codebase.

## The cybersecurity claims

[Google reports results](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/) across three stages of security work:

- **Finding vulnerabilities:** 88.2% on SEC-bench Pro.
- **Testing exploitability:** 54.8% on ExploitGym.
- **Patching:** 71.0% on CWE-bench.

Google says these external benchmark evaluations were conducted **without guardrails**. They should not be assumed to describe the behavior of the eventual public product, which will include safeguards, or to translate directly into success rates on real incidents. The company also says Carbon more than doubled Argon's success rate at generating secure code on an internal benchmark; it has not supplied enough detail in the announcement to independently assess that comparison.

These capabilities cut both ways: the ability to locate and validate vulnerabilities can help defenders, while also requiring care around misuse. Google says its initial access will go through a subset of defenders and expand to trusted partners in its Fairwind Program, while it gathers feedback and iterates on guardrails. It also says it is participating in the US government's voluntary pre-release testing process.

## Reliability and safety: encouraging claims, limited context

Google says Carbon builds on Argon's layered safety approach, including training, misuse defenses, prompt-injection resistance and monitoring. It reports that high-severity misalignment alerts in internal live coding environments fell by almost 80%, and that unsupported claims in internal testing fell by roughly half compared with Argon. Those are Google measurements; the announcement does not provide enough methodological detail to calculate how often a user will encounter either problem.

[Google also reports](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/) zero successful attacks in its evaluation on Gray Swan's indirect prompt-injection benchmark. Zero observed successes in a particular test is not proof that the model cannot be manipulated in other settings. Likewise, Google's reported zero successful cheating on Terminal-Bench 4.0 is a result for that evaluation, not a guarantee for all agent workflows.

## Access and price

Google says it will begin offering Carbon to a subset of cyber defenders this week, then gradually extend access to additional trusted partners. It intends to bring the model to developers, businesses and consumers as soon as possible, but has not announced a public launch date or detailed access limits.

For its eventual launch, [Google lists an introductory price](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/) of **$2 per million input tokens** and **$10 per million output tokens**, with cached input priced 95% below the input rate. After the introductory period, it says the price will be **$4 per million input tokens** and **$20 per million output tokens**. Google does not specify an end date for Carbon's introductory period in the announcement. These are planned prices, not evidence of current public API access.

For a point of reference, [Google made Argon broadly available](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-argon/) on October 13 through its developer and enterprise surfaces and to eligible paid users in its consumer products. Carbon's initial rollout is narrower.

## What to watch next

The most useful next disclosures would be an independently inspectable model and safety report, evaluation methods and comparisons, and actual product access. Developers will also need the API model identifier, context limits, rate limits and the dates governing introductory pricing before they can make a practical migration or cost decision. Until then, Carbon's headline numbers are a substantial set of Google-reported claims, with the public test still ahead.