Gemini 4 Carbon announced: benchmarks, price and access

Two weeks after Gemini 4 Argon, Google has announced its next model, Gemini 4 Carbon, and says it leads on cybersecurity. We read Google’s announcement, dated October 28 and written by Koray Kavukcuoglu. It is a vendor announcement: every number below is Google’s own, and Carbon is not yet generally available. We did not read a model card or run the model.

What did Google announce?

Carbon is described as Google’s “strongest model with state-of-the-art cybersecurity and long horizon capabilities”. Google says it ranks first on the recently updated version 5.0 of the Artificial Analysis Intelligence Index, at 59, and shows gains in coding, agentic tasks and cybersecurity defense. It says Argon “was launched two weeks ago” and that Carbon shows how quickly it is pushing the frontier, with its researchers and models “working hand in hand”.

Who can use it, and at what price?

• Access. Starting this week Carbon goes to a subset of cyber defenders, and access will be phased to more trusted partners through the Fairwind Program. Google says it is gathering feedback and iterating on guardrails before offering it to developers, businesses and consumers “as soon as possible”, and that it is taking part in the U.S. government’s voluntary process for pre-release model testing.
• Price. Carbon will launch at “the same introductory price” as Argon: $2 per million input tokens and $10 per million output tokens, with cached input tokens priced at 95% off the input price. The page carries a footnote on the introductory price that we did not read. Our Argon report found that Argon’s pricing page listed those rates through November 30 and $4 and $20 from December 1. Whether the same end date applies to Carbon is not stated in the text we read.

What does Google say it can do?

• Long tasks. Google says its teams have seen Carbon run for days on multi-step engineering problems such as root-causing and patching subtle bugs across large codebases and low-level systems. On WebDev Arena it says Carbon gained 89 Elo points to reach fourth place at 1768.
• Cybersecurity. Google reports, on external benchmarks evaluated without guardrails: 88.2% on SEC-bench Pro for finding vulnerabilities, 54.8% on ExploitGym for exploitation and 71.0% on Collinear’s CWE-bench for patching. On internal benchmarks it reports more than twice Argon’s success rate at generating secure code. It says the Chrome team found Carbon “more than 10X better” than “another lab’s large frontier model” at generating high-quality patches, and that Wiz sees it leading its web penetration-testing benchmarks. The comparison model in the Chrome claim is not named.

What does Google say about safety?

Carbon uses the same safety stack as Argon: training, real-time misuse defenses, prompt-injection resilience and chain-of-thought monitoring. Google reports, from its own tests:

• Successful cheating on Terminal-Bench 4.0 fell to 0.0%, with attempts at 0.4%.
• High-severity misalignment alerts in live internal coding dropped by almost 80%.
• Unsupported claims fell by about half against Argon.
• Overreach fell 26% and under-service 29%.
• On Gray Swan’s indirect prompt-injection benchmark the attack success rate is 0.0%, against 0.7% for Argon.

It also says Carbon surfaces source discrepancies to the user before acting instead of guessing.

What does Artificial Analysis say?

We checked Google’s headline claim against Artificial Analysis’s page for the model. It confirms the score: Gemini 4 Carbon (high) is first of 52 models on Intelligence Index v5.0 with 59, ahead of Claude Opus 5.5 at 58, GPT-6 Astra at 57, GPT-6.1 Sol at 55 and Claude Fable 5.1 at 49. The lead over Opus 5.5 is one point, which is our reading of the chart. The page also reports:

• Price and access. It lists $4.00 in and $20.00 out per million tokens with a 95% cache discount, not Google’s $2 and $10 introductory rates, and the chart notes “No publicly available API”. We cannot say why the prices differ; the standard rates after the introductory period are one possibility, and that is our guess.
• Cost per task. $7.76 per Intelligence Index task, against $10.04 for Opus 5.5 and $13.04 for Fable 5.1 on the same chart.
• Speed and verbosity. 92.4 output tokens per second, and 200M tokens generated on the index against a median of 64M, which the page calls very verbose.
• Specs. Text, image, speech and video input, text output and a 1M-token context window.

The Artificial Analysis score covers the index only. Google’s cyber benchmarks (SEC-bench Pro, ExploitGym, CWE-bench) are not part of it.

What is still unknown?

• Independent results beyond the index. We confirmed the Artificial Analysis score. The cyber, safety and WebDev Arena figures are as reported by Google; we did not check them against the benchmark authors.
• The model card. We did not find one in the announcement. The safety figures are summary claims.
• General availability. There is no date beyond “soon”, and the context window, model ID and rate limits are not in the text we read.
• What the government testing process involves and whether its results will be published.
• The end of the introductory price, as above.

What to keep in mind

These are our observations:

• A cyber-leading model released in stages is a pattern. Argon also went to Fairwind partners before its wider availability on October 13. OpenAI said this week that GPT-6.1 Terra meets the Critical cybersecurity threshold in its own framework. Staged release is how vendors are handling models strong at vulnerability work, though the criteria differ and are the vendors’ own.
• Few outsiders can test it today, so treat the cyber and safety tables as claims until independent evaluations arrive. The one-point index lead is real but small.
• Defenders and attackers read the same numbers. An exploitation score of 54.8% on a public benchmark is a statement about capability, not about what any given deployment is exposed to.
• Check price terms before planning around $2 and $10. Argon’s rates double in December, per its own pricing page.

Editorial review: October 28, 2026 (AI). Prepared from Google’s announcement page (read in full as page text), the Artificial Analysis model page (read as page text, first part), the Gemini 4 Argon pricing details from our own earlier report and the Hacker News thread’s title. We did not read the footnotes, a model card or the index methodology in full, did not run the model and did not contact Google. The page may have changed since our reading.