Google Broadens Access to Gemini 4 Carbon Model for Complex Reasoning
Google has broadened commercial access to its Gemini 4 Carbon artificial intelligence model, making the system available to software developers, enterprise customers, researchers and general consumers. Developed by Google DeepMind, Carbon is engineered for complex computations, software engineering workflows, cybersecurity auditing, and multi-stage autonomous tasks that require extended reasoning. The software architecture is designed to execute multi-step problem solving with continuous verification, complementing simpler neural networks within coordinated agent teams. The wider deployment follows a controlled preview that commenced on October 28 under Google's Fairwind initiative for authorized cybersecurity practitioners.
API Commercial Terms and Consumer Availability
The model is accessible through the Gemini API within Google AI Studio and enterprise commercial portals. During an introductory promotional window concluding November 30, developers can query the model at identical tariff rates to Argon, set at $2 per million input tokens and $10 per million output tokens. Standard rates will later rise to $4 per million input tokens and $20 per million output tokens, with a 95% discount for cached input tokens. For consumer tiers, Google is deploying Carbon across Google Antigravity, the Gemini mobile app, and AI Mode in Search for premium subscribers. Google DeepMind noted that while Argon remains preferred for daily software engineering, Carbon addresses demanding projects requiring deeper reasoning and repeated verification. Gemini 4 Carbon was added to LiteLLM's model price file on November 10, 2026, with availability across providers including Google AI Studio and Google Vertex AI. Gemini 4 Carbon follows Gemini 4 Argon, which became generally available in the Gemini API on October 13, 2026, according to SaaSCity. The model has a 1,048,576-token context window and returns up to 1,048,576 tokens per response, according to LiteLLM.
Benchmark Evaluations and Multi-Agent Collaboration
In competitive industry assessments, Carbon secured first place across CursorBench 4.0, Perplexity's WANDR, and Zapier's AutomationBench benchmark suites. On AutomationBench, Carbon high effort scored 53.88% at an operating expenditure of $2.12 per completed task, compared with an estimated $4.64 per task for Claude Fable 5.2 at maximum effort, which scored 51.45%. Artificial Analysis gave Carbon a score of 60 at high effort on its Intelligence Index, behind Claude Fable 5.2 at 65 and ahead of Claude Opus 5.5 at 59 and GPT-6 Astra at 57. On the Terminal-Bench 4.0 programming evaluation, Carbon attained 65.2%, above Claude Fable 5.2's best listed result of 61.6%. Alongside the model release, Google introduced a preview of the Swarm feature in Antigravity, enabling Carbon to direct multi-agent software development teams.
Specialized Capabilities in Finance and Cybersecurity
In institutional financial analysis, Carbon matched leading systems on Hebbia's benchmark and placed first among 23 model configurations evaluated by Balyasny Asset Management. Security testing under Project Fortify demonstrated 100 per cent recall in vulnerability detection alongside 95 per cent precision in incident triage. Security firm Wiz utilized the model to discover zero-day vulnerabilities across two software-as-a-service platforms in less than two hours. Independent testing by Gray Swan showed 0% successful prompt injections while the neural network blocked 99% of attempted security boundary bypasses.