Google unveils Gemini 4 Argon, which can write 1 million tokens in one response

Google DeepMind introduced its new frontier model, Gemini 4 Argon, on September 30, 2026. Built to sustain deep reasoning across long, multi-step workflows, the model targets real-world software engineering, enterprise knowledge work such as legal and finance, and cybersecurity defense. According to the announcement, it is not yet open to everyone: it is rolling out first to selected cyber defense teams in Google's Fairwind Program.

1 million tokens in a single response

Argon's standout feature is its output limit. According to Google, the model can write up to 1 million tokens in a single response, up from the previous limit of 64,000. The context window is how much a model can read at once, while the output limit is how much it can write in one response. The larger limit is meant for long jobs such as rewriting a large codebase or working through a multi-step problem in one go. The announcement does not state the size of the model's input context window.

Google says thousands of its employees already use the model internally and gives examples. According to the page, Argon agents found optimizations that freed a significant amount of memory across Google's data centers. The model has also been used to move C/C++ code to Rust, in projects ranging from re2 and libgav1 to the Zircon kernel of the Fuchsia operating system; Google notes that these rewrites are still being audited and tested before going into production. In libgav1, Google's open source video decoder, Argon agents are reported to have produced memory-safe code that runs faster and gives identical video output.

Benchmarks

Google's table compares Argon with GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.5. By the company's measurements, the model leads on knowledge work benchmarks such as the Vals Index, Vals Finance Agent v2 and Harvey's legal agent benchmark, and on DeepSWE, which measures long-horizon software engineering. Rivals come out ahead on some coding tests: GPT-6 Astra scores higher on FrontierSWE and Claude Opus 5.5 on Terminal-bench. On CWE-bench, which measures fixing security vulnerabilities, Argon shares first place with rival models.

First access goes to cyber defenders

Fairwind is Google DeepMind's controlled program that gives priority defenders early access. According to the program page, governments and national cyber authorities, critical infrastructure operators in healthcare, telecom, energy and finance, and core technology platforms can take part. Google says Argon is being released to these trusted defenders and to its own internal teams without cyber guardrails, so they can use the model's full defensive capabilities. Partners may use the model only for defensive work such as authorized threat simulation, reverse engineering and malware analysis. Creating malware is not allowed, access is limited to internal security teams, and it cannot be shared with others. Argon can be used on its own or inside CodeMender, Google's code security agent.

Wider availability will follow once feedback is gathered and safeguards are strengthened. Google says the model will open first to paid Gemini API customers and Google AI Ultra subscribers, but gives no date. Current prices are on Google's site.

Safety measures

Google says it is strengthening safeguards in four areas before broad release: refusing misuse requests, resisting prompt injection, monitoring the model's chain of thought and actions and stopping execution when needed, and using isolated sandboxes for high-risk training and evaluations. The company describes Argon as its most resilient model to prompt injection so far.

Doubts from inside Google

After the announcement, Bloomberg reported, in a story by Julia Love and Davey Alba, that Google is grappling with internal skepticism about the model. Citing people with direct access to the effort, the report says some Google employees find that Gemini 4 does well on common benchmarks but less well when put to real work, particularly on certain coding tasks. Google told Bloomberg it would be "inaccurate" to say the model underperforms in coding. According to the report, another employee said there is broad internal consensus that the model is at the frontier. Bloomberg also recalls that Google dropped Gemini 3.5 Pro, which it had promised for June.

All the published benchmarks were run by Google, and the model is currently open only to a limited group. Only wide use will show the gap between benchmark results and daily use, how coherent such long outputs stay from start to finish, and what they cost. To see how the model really performs, it is worth following evaluations from independent researchers.

Sources: Google: Gemini 4 Argon, Google DeepMind: Fairwind Program, Google DeepMind: Frontier Safety Framework, Bloomberg: Google grapples with employee skepticism about new Gemini model.