Google unveiled Gemini 4 Carbon on Wednesday, calling it the company's most capable model to date and placing it first on version 5.0 of the Artificial Analysis Intelligence Index with a score of 59. That puts Carbon one point ahead of Anthropic's Claude Opus 5.5 at 58 and two ahead of OpenAI's GPT-6 Astra at 57. The previous Gemini flagship, Gemini 4 Argon, sits at 50.
Carbon lands just two weeks after Argon, which was first announced on September 30 and rolled out more broadly on October 13. Google DeepMind SVP and Google Chief AI Architect Koray Kavukcuoglu wrote the launch post, which frames Carbon around two strengths: cybersecurity and long-horizon agentic work such as multi-step engineering tasks and long-running coding sessions.
“Following Argon's release just two weeks ago, Carbon shows the pace at which we're pushing the frontier”— Sundar Pichai, CEO of Google
On security, Google splits the work into finding, exploiting and patching vulnerabilities. Carbon scored 88.2% on SEC-bench Pro for discovery, up from Argon's 83.4%, 54.8% on ExploitGym for exploitation, up from 42.0%, and 71.0% on CWE-bench for patching, up from 68.0%. Google says those external security evaluations ran without guardrails, so restricted customer deployments may score differently.
Key facts
- 01Gemini 4 Carbon scored 59 on Artificial Analysis Intelligence Index v5.0, ahead of Claude Opus 5.5 (58) and GPT-6 Astra (57).
- 02Google reports 88.2% on SEC-bench Pro, 54.8% on ExploitGym and 71.0% on CWE-bench, all up from Gemini 4 Argon.
- 03Introductory pricing is $2 per million input tokens and $10 per million output tokens, rising later to $4 and $20.
- 04Access starts this week with a subset of cyber defenders through Google's Fairwind Program; no date for a broad release.
- 05Carbon is part of the US government's voluntary pre-release model testing process.
Google also cited partner results. The Chrome team reported that Carbon produced high-quality patches more than 10 times better than another lab's large frontier model, which Google did not name, and Wiz reported leading results on web penetration-testing benchmarks. In internal tests, Google says Carbon more than doubled Argon's success rate at generating secure code.
The general agentic numbers moved up too. Carbon scored 65.2% on Terminal-Bench 4.0 versus Argon's 57.1%, 77.4% on Agents' Last Exam versus 69.0%, and 96.2% on the GraphWalks long-context test, ahead of GPT-6 Astra at 71.8% and Opus 5.5 at 66.8%. On WebDev Arena it gained 89 Elo points to 1769, good for fourth place. In Andon Labs' Vending-Bench 2 business simulation, Carbon finished with $16,484, ahead of GPT-6 Astra's $15,515.
“Internally, our teams have seen Carbon run for days”— Koray Kavukcuoglu, SVP, Google DeepMind and Chief AI Architect, Google
Carbon does not lead every chart. Google's own comparison table shows it with the top score in 15 of 17 rows. GPT-6 Astra scored 65.5% and Opus 5.5 scored 62.3% on FrontierSWE 2 against Carbon's 61.2%, and Opus 5.5 led the OSWorld offline subset at 87.2% to Carbon's 81.5%. On the AA-Omniscience knowledge test Carbon posted the top score of 54 and answered 58% of questions correctly, though its measured hallucination rate rose from Argon's 4% to 11%.
Google paired the capability gains with reliability numbers. The company says Carbon cut successful cheating on Terminal-Bench 4.0 to 0.0%, with cheating attempts at 0.4%, and recorded a 0.0% attack success rate on Gray Swan's indirect prompt-injection benchmark, versus 0.7% for Argon. Internally, Google reports high-severity misalignment alerts in live coding fell by almost 80%, unsupported claims dropped by about half, overreach fell 26% and under-service fell 29%.
Pricing matches Argon's. Carbon costs $2 per million input tokens and $10 per million output tokens during an introductory period, with cached input tokens discounted 95%. After that, rates rise to $4 per million input tokens and $20 per million output tokens. Google has set November 30 as the end of Argon's introductory pricing but has not given an end date for Carbon's.
Access is the catch. Starting this week, Google is giving Carbon to a subset of cyber defenders through its Fairwind Program, then expanding gradually to trusted partners. Google says it will bring Carbon to developers, businesses and consumers across its products as soon as possible, after gathering feedback and refining guardrails, but gave no date. Google has not published a technical report, parameter count or architecture details.
Carbon is also part of the US government's voluntary pre-release model testing process, according to Google. For now, Gemini 4 Argon remains the newest Google model that most developers and customers can use.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
