Thursday, 29 October 2026

Google’s Gemini 4 Carbon tops AI index, starts with defenders

Google has released Gemini 4 Carbon to a small group of cyber defenders first, with wider access for developers and consumers still to come.

Key points

  • Gemini 4 Carbon ranks first on Artificial Analysis Intelligence Index v5.0 with a score of 59.
  • Access begins with cyber defenders; Google gives no date for general availability.
  • Introductory API pricing is $2 input and $10 output per million tokens.

Google has released Gemini 4 Carbon, its newest frontier model, but only to a small group of cybersecurity defenders for now. Developers, businesses and ordinary users will have to wait; Google says broader access is “coming soon” and has not given a date.

Carbon arrives roughly two weeks after Gemini 4 Argon, the model that reached everyday users earlier this month. Google is pitching the new one at long, multi-step coding jobs and at security work, where it says the gains are largest.

A score of 59 and a security focus

In its launch post, Google says Carbon puts Carbon at number one on the Artificial Analysis Intelligence Index (v5.0) with 59 points. On WebDev Arena it sits fourth at 1768 Elo, up 89 points. Koray Kavukcuoglu, senior vice president at Google DeepMind, calls it the company’s “strongest model with state-of-the-art cybersecurity and long horizon capabilities”.

Security benchmark What it tests Carbon
SEC-bench Pro Finding vulnerabilities 88.2%
ExploitGym Exploiting them 54.8%
CWE-bench Patching them 71.0%

Google says those runs were external and done without guardrails. Internally, it reports more than double Argon’s success rate at writing secure code, and says the Chrome team found Carbon produced over 10 times more high-quality patches than another lab’s large model. These are Google’s own figures, not independent tests.

Fewer shortcuts, and a rollout under review

Google also publishes behaviour figures. Carbon did not cheat on the Terminal-Bench 4.0 coding test, and the Gray Swan prompt-injection attacks never got through (Argon was hit 0.7% of the time). Google says its own coding teams saw serious misalignment flags fall by nearly 80%. The full detail sits in a separate safety report.

This week’s rollout covers a limited set of defenders, and Google plans to add trusted partners gradually through its Fairwind program. The company is also putting the model through the US government’s voluntary testing before release.

Pricing once it opens up

Developers will pay an introductory $2 for every million tokens they send in and $10 for every million they get back, the same as Argon’s launch rates, and cached input costs 95% less. Once the introductory period ends, those rates double to $4 and $20. We covered the company’s other recent model work in our Genie 4 report. Google has not said whether or when Carbon will reach Gemini app users, so we will update this post when it does.

TPRO Team

The Tech Prolonged newsroom desk. Stories under this byline cover AI, the web and general tech, and are reviewed by our editors before publishing.