Google launches Gemini 4 Carbon with top AI Index ranking

Google introduced Gemini 4 Carbon on Oct. 28, 2026, roughly two weeks after Argon reached Google AI Pro and Ultra subscribers in the Gemini app on Oct. 13. Koray Kavukcuoglu, an SVP at Google DeepMind, described Carbon as the company’s strongest model in an Oct. 28 blog post. Kavukcuoglu leads Google DeepMind after Demis Hassabis stepped aside.

Carbon is available first to vetted cyber defenders through Google’s Fairwind Program, with trusted defenders receiving access first this week. Broader Fairwind Program expansion will follow, though Google has not provided dates for when Carbon will reach developers, businesses, and consumers or for the price increase.

Sundar Pichai framed Carbon’s rapid release as proof that Google intends to push the frontier quickly. Pichai said Carbon “ranks #1 on the recently updated v5.0 of Artificial Analysis’ Intelligence Index at 59”.

Google said cybersecurity and long-horizon tasks are Carbon’s biggest strengths. According to Google, Carbon can tackle hard engineering problems over periods lasting days.

Carbon ranks number one on Artificial Analysis Intelligence Index version 5.0 with a score of 59. Claude Opus 5.5 scored 58 and GPT-6 Astra scored 57 on the same index. Artificial Analysis refreshed its Intelligence Index to version 5.0 around the time Carbon launched. The one-point lead is thin enough that the next release from OpenAI or Anthropic could erase it.

Google’s lead on the index may be short-lived because rivals ship new models quickly. When checked on Wednesday, the Artificial Analysis public leaderboard had not yet added Carbon. Intelligence Index version 5.0 scores cannot be directly compared with the previous version, where Argon High scored 53.

Google, OpenAI, and Anthropic have each claimed a new lead within weeks of the last, intensifying competition over benchmark supremacy. Whether the lead holds is likely to be short-lived, given the current cadence of releases from all three labs. All three AI labs released new models this month.

Anthropic released Claude Opus 5.5 on Sept. 22, and OpenAI shipped GPT-6 Astra earlier that month. Google announced Gemini 4 Argon on Sept. 30, and Argon reached Google AI Pro and Ultra subscribers on Oct. 13. Argon’s debut led on 12 of 18 comparisons against GPT-6 Astra and Claude Opus 5.5. Carbon’s biggest gains over Argon are in long-context retrieval, science reasoning, and cybersecurity defense.

Carbon achieved the highest result on 15 of 17 benchmarks in Google’s comparison table. Google’s comparison table included Argon, GPT-6 Astra, Claude Opus 5.5, and Claude Fable 5.1. Google’s own benchmark table provides most of the head-to-head figures. Company-reported scores for rival models did not consistently agree with independent testing outcomes.

Carbon led on most disclosed benchmarks, with clear losses only on OSWorld-2.1 and FrontierSWE 2.

GPT-6 Astra leads on FrontierSWE 2, a software engineering test. Claude Opus 5.5 leads on OSWorld-2.1, a test of computer operation ability. On FrontierSWE 2, GPT-6 Astra scored 65.5 percent, Claude Opus 5.5 scored 62.3 percent, and Carbon scored 61.2 percent.

Carbon’s Terminal-Bench 4.0 score improved by roughly eight points versus Argon and surpassed Opus 5.5. Terminal-Bench 4.0 gauges an AI model’s ability to finish real technical tasks using command-line input alone. This benchmark ranks among the most difficult agentic evaluations for large language models. For Terminal-Bench 4.0, Google uses scores run by Artificial Analysis instead of its own results, unlike with Argon.

OfficeChai reported that Carbon scored 65.2 percent on Terminal-Bench 4.0 run by Artificial Analysis, ahead of Opus 5.5 at 59.6 percent and Argon at 57.1 percent. Google said Carbon’s web development skills improved enough to reach fourth place on WebDev Arena. 9To5Google reported that Carbon jumps 89 elo points to number 4 with 1769 on WebDev Arena.

9To5Google said Google described a firm lead for Carbon in three cybersecurity benchmarks: SEC-bench Pro for vulnerability discovery, ExploitGym for exploitation, and CWE-bench for patching. OfficeChai reported Carbon scored 88.2 percent on SEC-bench Pro and 71.0 percent on CWE-bench. Yahoo Tech reported that on ExploitGym, Carbon scored 54.8 percent, GPT-6 Astra scored 42.4 percent, and Claude Opus 5.5 scored 34.5 percent.

At launch, Carbon charges $2 for each million input tokens and $10 for each million output tokens. Cached input tokens cost 95 percent less than the standard input price. After the introductory period, Carbon will cost $4 per million input tokens and $20 per million output tokens. Google has not provided dates for broader availability or the price increase.

Claude Opus 5.5 charges $4 for each million input tokens and $20 for each million output tokens. GPT-6 Astra charges $10 for each million input tokens and $50 for each million output tokens. OpenAI cut GPT-6.1 Terra prices to $1 and $5 on October 26.

Alphabet’s stock increased roughly 1.5 percent the day Carbon claimed the top index position. Anthropic is preparing for its planned stock listing.

Carbon relies on Argon’s safety system, which monitors the model’s reasoning, plus alignment work. According to Google’s own testing, Carbon behaves better than its predecessor. All safety figures come from Google’s own testing.

Carbon did not succeed in cheating on Terminal-Bench 4.0. Carbon’s cheating attempts on Terminal-Bench 4.0 occurred in just 0.4 percent of cases. High-severity misalignment alerts during internal coding work fell by nearly 80 percent compared with Argon. Carbon generates roughly 50 percent fewer unsupported claims than Argon.

Carbon stopped all attack attempts in Gray Swan’s prompt injection test, while Argon’s attack success rate on that benchmark was 0.7 percent.

Google is participating in the US government’s voluntary testing program for models before release. Outside cybersecurity tests ran with safety filters switched off. The staged release puts safety first. The staged release reflects industry-wide caution around models with strong offensive cybersecurity capabilities.

Cybersecurity incidents rattled the industry throughout the month. Carbon and Meta’s Muse Wonder Full have no publicly available API on the leaderboard.

Carbon scored 96.2 percent on GraphWalks, a long-context test spanning 256k to 1M tokens. That result is more than 12 points above Argon and roughly 24 points above GPT-6 Astra. Carbon scored 53.9 percent on AutomationBench, a knowledge work test. Carbon finished with $16,484 on Vending-Bench 2, a simulated business test.