Google releases Gemini 4 Carbon AI model focused on cybersecurity and coding

On October 28, 2026, Google unveiled Gemini 4 Carbon, a frontier AI model tailored for cybersecurity defense and agentic coding, released two weeks after Gemini 4 Argon. Gemini 4 Carbon topped the Artificial Analysis Intelligence Index v5.0 with a score of 59, leading competitors on cybersecurity tests like SEC-bench Pro (88.2%) and CWE-bench (71.0%). Google restricted initial access to vetted cyber defenders via its Fairwind Program at introductory rates of $2 per million input tokens.

Launch and design purpose

• ▪Alphabet's Google DeepMind launched Gemini 4 Carbon, built for long-horizon agentic coding and cybersecurity defense
• ▪Google DeepMind designed Gemini 4 Carbon for complex engineering work, allowing internal teams to run single tasks across large codebases for days
• ▪Google launched Gemini 4 Carbon two weeks after introducing its predecessor, Gemini 4 Argon

Cybersecurity and patching benchmarks

• ▪Google said Gemini 4 Carbon scored 71.0% on Collinear's CWE-bench for software patching, higher than the 68.0% scored by both Gemini 4 Argon and GPT-6 Astra
• ▪Internal benchmarks showed Google's Gemini 4 Carbon generated secure code at more than twice the success rate of Gemini 4 Argon, with Google's Chrome team reporting over 10 times better high-quality patching performance than another lab's model
• ▪On the ExploitGym benchmark, which Google evaluated using its Antigravity harness with mitigations off, Gemini 4 Carbon scored 54.8% for vulnerability exploitation compared to 42.4% for GPT-6 Astra and 34.5% for Claude Opus 5.5

Coding and long-context benchmarks

• ▪Google said that on the Terminal-Bench 4.0 agentic coding benchmark, Gemini 4 Carbon scored 65.2%, placing it ahead of Claude Opus 5.5's 59.6%
• ▪Google said Gemini 4 Carbon scored 9.6% on SWE-sweep, narrowly leading GPT-6 Astra's 9.5%, while scoring 80.0% on RiemannBench and 77.4% on Agents' Last Exam v1.1
• ▪Google's Gemini 4 Carbon achieved a 96.2% score on the GraphWalks long-context benchmark spanning 256K to 1M tokens, compared to 84.2% for Gemini 4 Argon and 71.8% for GPT-6 Astra
• ▪Gemini 4 Carbon gained 89 Elo points on WebDev Arena to rank fourth with a score of 1769 for full-stack web development

Fairwind Program early access

• ▪Google's Fairwind Program prioritizes early Gemini 4 Carbon access for governments, critical infrastructure operators, and core tech platforms before opening to general developers
• ▪Google limited initial access to Gemini 4 Carbon to a subset of cybersecurity defenders through its Fairwind Program during the week of October 28, 2026, while participating in U.S. government voluntary pre-release testing

Token pricing and discounts

• ▪Google offers a 95% discount on cached input tokens for Gemini 4 Carbon
• ▪Google said Gemini 4 Carbon carries introductory pricing of $2 per million input tokens and $10 per million output tokens, which will later increase to $4 and $20 per million tokens respectively

Safety and alignment improvements

• ▪Google measured a roughly two-fold reduction in unsupported claims and alignment gains including a 26% drop in overreach and a 29% drop in under-service compared to Gemini 4 Argon
• ▪Attack success against Gemini 4 Carbon on Gray Swan's Indirect Prompt Injection benchmark fell to 0.0%, down from 0.7% for Gemini 4 Argon, Google said
• ▪Google reported that successful cheating by Gemini 4 Carbon on Terminal-Bench 4.0 dropped to 0.0%, while high-severity misalignment alerts in live internal coding fell by almost 80%

Competitor advantages on specific benchmarks

• ▪On the FrontierSWE benchmark, OpenAI's GPT-6 Astra led with 65.5% accuracy compared to Gemini 4 Carbon's 61.2%, Google said
• ▪Google said Anthropic's Claude Opus 5.5 outperformed Gemini 4 Carbon on the OSWorld-2.1 computer use benchmark, scoring 87.2% against Carbon's 81.5%