Google releases Gemini 4 Carbon for long-horizon agents
Google’s new frontier model targets extended coding, scientific, cybersecurity and knowledge-work tasks with a 1-million-token context.
Google DeepMind has released Gemini 4 Carbon, a frontier model aimed at tasks that require sustained planning, tool use and verification rather than a single answer. The model is available through Google’s consumer and developer channels, and has also begun appearing in products including GitHub Copilot, Cursor, Devin, Perplexity and Databricks.
What changed
DeepMind describes Carbon as its strongest model for long-horizon agentic work across complex knowledge tasks, software engineering, cybersecurity and scientific reasoning. Its model card lists text, image, audio and video inputs, with a combined input-output context window of up to one million tokens. That capacity is intended for large codebases, lengthy technical files and multi-stage investigations that would otherwise require repeated context handoffs.
Google is also emphasizing extended verification. The company says Carbon can spend longer exploring a problem, construct intermediate artifacts and check its work before returning a result. Early third-party distribution suggests the launch is as much about becoming the default reasoning layer inside developer and enterprise tools as it is about winning a standalone chatbot benchmark.
The rollout is unusually broad for a new frontier model. Availability through coding environments and agent products gives developers immediate ways to compare its performance against existing models, while discounted introductory pricing in some tools lowers the cost of experimentation. Those integrations may matter more than launch-day leaderboard positions because they expose Carbon to real workflows with persistent state and external actions.
Why it matters
The release moves competition toward models that can remain useful across long tasks, not merely answer difficult prompts. It also raises the operational bar: a one-million-token context window is valuable only if agents can retrieve the right material, control costs and avoid compounding small errors over hours of work. Carbon’s real significance will be measured by whether developers keep it in production after the promotional pricing ends and whether its claimed long-horizon strengths survive messy, tool-connected environments.