Google’s Gemini 4 Carbon Tops AI Benchmark

Google DeepMind announced its flagship Gemini 4 Carbon on the 28th, local time, in an official blog post. Carbon scored 59 on the Artificial Analysis Intelligence Index (AAII) at the High reasoning setting, placing first among 52 models. Claude Opus 5.5 scored 58, GPT-6 Astra 57, and its predecessor Gemini 4 Argon 50. AAII combines nine benchmarks covering coding, reasoning, scientific problem-solving, and professional tasks.

Google took first place in the assessment for the first time in about 11 months. Gemini 3 Pro, released on November 18 last year, had led GPT-5.1 by 3 points; the lead changed after OpenAI and Anthropic released models. Carbon was announced about two weeks after Argon, first shown in late September, was officially released to developers on October 13. Koray Kavukcuoglu, Google DeepMind’s senior vice president, said Carbon is the company’s strongest model for cybersecurity and long-running tasks.

Google said Carbon is designed to plan multistep work, revise failed approaches, and continue complex tasks autonomously. Internal researchers observed it performing engineering work over several days, tracing and fixing complex causes of errors in large codebases. Out-of-scope actions fell 26%, while failures to complete required tasks fell 29%. Carbon scored 59 on the Coding Agent Index, ranking second to Claude Opus 5.5 at 62. It also placed second on Terminal-Bench 4.0 with 64% and FrontierCode v1.1 with 51%.

In cybersecurity evaluations, Carbon scored 88.2% on SEC-Bench Pro, which tests finding vulnerabilities in real software; 54.8% on ExploitGym, which tests whether vulnerabilities can be exploited; and 71.0% on CWE-Bench, which tests vulnerability fixes. Google said its rate of successfully generating secure code was more than twice Argon’s. In an evaluation by Chrome’s development team, Carbon was more than 10 times as capable of producing high-quality security patches as large models from other AI labs. Internal safety evaluations found serious alignment-violation warnings down about 80% and unsupported factual claims down about 50%. The success rate of prompt-injection attacks fell from Argon’s 0.7% to 0%.

On Vals AI’s Vals Index v2.2 assessment of professional tasks in finance, law, coding, and tax, Carbon ranked second among 29 models with 70.36%. Claude Sonnet 5.5 led with 70.85%, while Claude Opus 5.5 scored 70.18%. Carbon ranked first on the overall finance index and legal research benchmark, second in financial modeling, and third on the tax-agent task. However, it used 64% more output tokens than Argon, took more than twice as long on finance and tax-agent tasks, and needed over 50 minutes per legal-research task. It was faster and more accurate than Argon in financial modeling and terminal coding. Its cost per task across the full assessment was $10.51, below Claude Sonnet 5.5’s $18.70.

In Artificial Analysis’s AA-Benchcase, Carbon scored 1,609, 126 points above Argon’s 1,483, but behind Kimi K3.1 Pro at 1,678, Muse Wonder at 1,659, and Grok 4.7 at 1,638. Carbon ranked first on AA-Omniscience with 54, 6 points above Claude Opus 5.5. Its accuracy rose from Argon’s 48% to 58%, while its hallucination rate increased from 4% to 11%. In the same evaluation, Claude Opus 5.5 and GPT-6 Astra each had a 50% hallucination rate; Artificial Analysis rated Carbon’s rate as the second-lowest after Argon’s.

Carbon supports a context window of up to 1 million tokens and accepts text, image, audio, and video inputs, with text-only output. Google has not disclosed its parameter scale or specific training structure. Initial API prices are $2 per 1 million input tokens and $10 per 1 million output tokens. After the discount ends, the prices will double to $4 and $20, respectively; Google has not said when the discount will end. At list price, Carbon costs $7.76 per task, about 23% less than Claude Opus 5.5 at $10.04 but more than GPT-6 Astra at $6.25. Carbon generated about 200 million output tokens in the evaluation, more than three times the comparison models’ median of 64 million.

Starting this week, Google will provide Carbon to selected cybersecurity experts and plans to expand access to vetted partners through its Fairwind program. The company is also participating in US government evaluations of models before release. Google has not announced a general release date for developers, businesses, or consumers. It plans to validate safeguards using early-user feedback before widening access.

Google DeepMind announced its flagship Gemini 4 Carbon on the 28th, local time, in an official blog post. Carbon scored 59 on the Artificial Analysis Intelligence Index (AAII) at the High reasoning setting, placing first among 52 models. Claude Opus 5.5 scored 58, GPT-6 Astra 57, and its predecessor Gemini 4 Argon 50. AAII combines nine benchmarks covering coding, reasoning, scientific problem-solving, and professional tasks.

Google took first place in the assessment for the first time in about 11 months. Gemini 3 Pro, released on November 18 last year, had led GPT-5.1 by 3 points; the lead changed after OpenAI and Anthropic released models. Carbon was announced about two weeks after Argon, first shown in late September, was officially released to developers on October 13. Koray Kavukcuoglu, Google DeepMind’s senior vice president, said Carbon is the company’s strongest model for cybersecurity and long-running tasks.

Google said Carbon is designed to plan multistep work, revise failed approaches, and continue complex tasks autonomously. Internal researchers observed it performing engineering work over several days, tracing and fixing complex causes of errors in large codebases. Out-of-scope actions fell 26%, while failures to complete required tasks fell 29%. Carbon scored 59 on the Coding Agent Index, ranking second to Claude Opus 5.5 at 62. It also placed second on Terminal-Bench 4.0 with 64% and FrontierCode v1.1 with 51%.

In cybersecurity evaluations, Carbon scored 88.2% on SEC-Bench Pro, which tests finding vulnerabilities in real software; 54.8% on ExploitGym, which tests whether vulnerabilities can be exploited; and 71.0% on CWE-Bench, which tests vulnerability fixes. Google said its rate of successfully generating secure code was more than twice Argon’s. In an evaluation by Chrome’s development team, Carbon was more than 10 times as capable of producing high-quality security patches as large models from other AI labs. Internal safety evaluations found serious alignment-violation warnings down about 80% and unsupported factual claims down about 50%. The success rate of prompt-injection attacks fell from Argon’s 0.7% to 0%.

On Vals AI’s Vals Index v2.2 assessment of professional tasks in finance, law, coding, and tax, Carbon ranked second among 29 models with 70.36%. Claude Sonnet 5.5 led with 70.85%, while Claude Opus 5.5 scored 70.18%. Carbon ranked first on the overall finance index and legal research benchmark, second in financial modeling, and third on the tax-agent task. However, it used 64% more output tokens than Argon, took more than twice as long on finance and tax-agent tasks, and needed over 50 minutes per legal-research task. It was faster and more accurate than Argon in financial modeling and terminal coding. Its cost per task across the full assessment was $10.51, below Claude Sonnet 5.5’s $18.70.

In Artificial Analysis’s AA-Benchcase, Carbon scored 1,609, 126 points above Argon’s 1,483, but behind Kimi K3.1 Pro at 1,678, Muse Wonder at 1,659, and Grok 4.7 at 1,638. Carbon ranked first on AA-Omniscience with 54, 6 points above Claude Opus 5.5. Its accuracy rose from Argon’s 48% to 58%, while its hallucination rate increased from 4% to 11%. In the same evaluation, Claude Opus 5.5 and GPT-6 Astra each had a 50% hallucination rate; Artificial Analysis rated Carbon’s rate as the second-lowest after Argon’s.

Carbon supports a context window of up to 1 million tokens and accepts text, image, audio, and video inputs, with text-only output. Google has not disclosed its parameter scale or specific training structure. Initial API prices are $2 per 1 million input tokens and $10 per 1 million output tokens. After the discount ends, the prices will double to $4 and $20, respectively; Google has not said when the discount will end. At list price, Carbon costs $7.76 per task, about 23% less than Claude Opus 5.5 at $10.04 but more than GPT-6 Astra at $6.25. Carbon generated about 200 million output tokens in the evaluation, more than three times the comparison models’ median of 64 million.

Starting this week, Google will provide Carbon to selected cybersecurity experts and plans to expand access to vetted partners through its Fairwind program. The company is also participating in US government evaluations of models before release. Google has not announced a general release date for developers, businesses, or consumers. It plans to validate safeguards using early-user feedback before widening access.