Gemini 4 Carbon is priced like Argon, and DeepMind says it leads the new index at 59

What happened

Google DeepMind published Gemini 4 Carbon on October 28, 2026. DeepMind says it ranks first on Artificial Analysis Intelligence Index v5.0 at 59, and that the introductory price matches Argon: $2 per million input tokens and $10 per million output tokens, with cached input at 95 percent off the input price. After that introductory period, the price is $4 and $20. The post does not date the end of the introduction. DeepMind says the model is in the U.S. government's voluntary pre-release testing. This week it goes to a subset of cyber defenders, then to more trusted partners through a program DeepMind calls Fairwind. It is not yet on the developer, business, or consumer surfaces. On cybersecurity tests run without guardrails, DeepMind reports 88.2 percent on SEC-bench Pro, 54.8 percent on ExploitGym, and 71.0 percent on Collinear's CWE-bench. Those are scores, not a method. An internal secure-code check is more than twice Argon's success rate. DeepMind says the Chrome team found Carbon more than 10 times better than another lab's frontier model at writing patches, and does not name that model. On WebDev Arena, DeepMind says Carbon gained 89 Elo and sits at 1,769, fourth place. Successful cheating on Terminal-Bench 4.0 fell to 0.0 percent, with attempts at 0.4 percent. DeepMind does not print Carbon's Terminal-Bench accuracy. On Gray Swan's indirect prompt-injection benchmark, DeepMind says the attack success rate is 0.0 percent, against 0.7 percent for Argon.

Why it matters

Argon opened to developers two weeks ago at this price. Carbon is the next model at the same introductory price, and the public API does not have it. A 59 on the new index, if it holds, would sit one point above the Opus 5.5 score Artificial Analysis printed on October 27. Until a developer can call Carbon, that 59 is a claim about a model you cannot use.

DeepMind says high-severity misalignment alerts in its own coding tools fell by almost 80 percent, and that unsupported claims fell by about half versus Argon. Both are internal.

Overreach fell 26 percent and under-service fell 29 percent, on DeepMind's account. The post does not define the yardstick.

Wiz says Carbon leads on its web security tests and does not print the score.