Google DeepMind announced Gemini 4 Carbon on October 28, 2026, a frontier model it says ranks first on version 5.0 of the Artificial Analysis Intelligence Index with a score of 59. Carbon is not yet available to the public: starting this week it goes to a subset of cyber defenders, with more trusted partners added gradually through Google’s Fairwind Program. The model arrives only two weeks after Gemini 4 Argon, and its headline gains are in cybersecurity and long-running engineering work.
What Google announced
Google calls Carbon its strongest model, with “state-of-the-art cybersecurity and long horizon capabilities.” It builds on Gemini 4 Argon, which shipped on October 13. Google says engineers inside the company have watched Carbon work for days on hard tasks, such as tracing and patching subtle bugs across large codebases and low-level systems, without giving up at dead ends.
On coding, Google says Carbon gained 89 Elo points on WebDev Arena and now sits fourth with 1768. The company also cites better output quality for full-stack web development and multimodal understanding. Pricing is set to match Argon’s introductory rate: $2 per million input tokens and $10 per million output tokens, with cached input tokens 95% cheaper than the standard input price. Ars Technica reports that Google would not give an end date for that introductory rate, while Argon’s discounted pricing runs through the end of November.
The cybersecurity numbers
Google reports results on three external benchmarks, all measured without safety guardrails:
- SEC-bench Pro, which tests finding real vulnerabilities: 88.2%.
- ExploitGym, which tests whether a model can exploit a flaw: 54.8%.
- CWE-bench from Collinear, which tests patching at the root cause: 71.0%.
Google says Carbon more than doubled Argon’s success rate at generating secure code on internal benchmarks, and that the Chrome team saw it produce high-quality patches more than 10 times better than another lab’s large frontier model. Cloud security company Wiz, Google adds, found Carbon leading on web penetration-testing benchmarks. These are Google’s figures, and several come from internal tests we cannot check. Ars Technica notes that the gaps to rivals such as GPT-6 Astra, Claude Opus 5.5 and Claude Fable 5.1 are a few points on most tests, and that every frontier model currently scores in the same neighborhood.
Safety claims
Google devotes much of the announcement to alignment. It says Carbon cut successful cheating on Terminal-Bench 4.0 to 0.0%, with attempted cheating at 0.4%, while improving accuracy. In live coding environments with its own teams, high-severity misalignment alerts fell by almost 80%, and unsupported claims dropped by roughly half compared with Argon. The company also reports 26% less overreach and 29% less under-service when the model follows multi-step plans, and a 0.0% attack success rate on Gray Swan’s indirect prompt-injection benchmark, down from Argon’s 0.7%. Ars lists rivals on that same test at 1% for Opus 5.5 and 8.5% for GPT-6 Astra. All of this is self-reported, and Google has not published the full evaluation details in the announcement.
Why access is limited
Google says it is “actively engaged in the U.S. government’s voluntary process for pre-release model testing” and wants to tune guardrails before opening Carbon to developers, businesses and consumers “as soon as possible.” It gave no date. The approach mirrors how the company handled Argon, and it reflects a broader pattern: models strong at finding and exploiting software flaws are being released first to defenders. Whether that stays a norm may depend on competitors such as Mistral, which released open weights for a cyber-focused model the next day, as we report in our coverage of Mistral Large 4.
Critics have a fair question about independence. Ars points out that Google presents Carbon as a new model, but that it could be a post-training upgrade of Argon, since the two share a lot and Carbon edges Argon by slim margins on every test shown. Google has not disclosed training details.
What to watch next
- The date Carbon reaches the Gemini API, Google AI Studio and paid Gemini plans, and whether the $2 and $10 introductory price holds.
- Independent results on SEC-bench Pro, ExploitGym and CWE-bench from outside the vendor.
- Whether the Artificial Analysis index still lists Carbon at 59 once rivals ship updates.
- Outcomes of the US government’s voluntary pre-release testing, and whether Google shares findings.
- How quickly the Fairwind Program expands beyond the first group of defenders.
Key takeaways
- Gemini 4 Carbon is Google’s new top model, claimed #1 on the Artificial Analysis Intelligence Index v5.0 at 59.
- Access begins with cyber defenders through the Fairwind Program; no public release date was given.
- Google reports 88.2% on SEC-bench Pro, 54.8% on ExploitGym and 71.0% on CWE-bench, all vendor-reported.
- Introductory API pricing matches Argon at $2 and $10 per million tokens.
- Safety claims, including near-zero cheating and prompt-injection success, have not yet been independently verified.
Frequently asked questions
What is Gemini 4 Carbon?
Gemini 4 Carbon is Google DeepMind’s newest frontier model, announced on October 28, 2026. Google says it is its strongest model for cybersecurity and long-horizon engineering tasks.
Can I use Gemini 4 Carbon today?
Not yet. Google is giving access first to a subset of cyber defenders and plans to add trusted partners through its Fairwind Program before a broader release.
How much will Gemini 4 Carbon cost?
Google lists an introductory price of $2 per million input tokens and $10 per million output tokens, the same as Gemini 4 Argon, with cached input 95% cheaper. The company has not said when the introductory rate ends.
