Google DeepMind has announced Gemini 4 Argon, a frontier model designed for lengthy professional tasks, with a one-million-token output limit and an initial focus on trusted cyber defenders. The launch pairs ambitious performance claims with a phased rollout, an introductory API price and striking examples of AI doing consequential work inside Google.
Google DeepMind Watch analysis
What happened
Argon is rolling out first to trusted cyber defenders through Google’s Fairwind programme. Google says it will widen access after testing and iterating on safeguards, with paid API customers and Google AI Ultra subscribers among the intended early users. Its introductory price is $2 per million input tokens and $10 per million output tokens; cached input tokens are priced at 95% off the input rate.
Google says Argon can handle long, multi-step work across software engineering, finance, legal tasks and cyber defence. The company highlights a one-million-token output limit, up from 64,000 tokens, and says Argon scored 77.9% on the DeepSWE v1.1 software-engineering benchmark and 51.3% on Zapier’s AutomationBench. These are Google’s reported results, not independent evaluations.
Our top picks
- A much longer runway for complex work
The one-million-token output limit is designed to let Argon sustain longer tasks and generate extensive responses in one trajectory. - A substantial code-migration example
Google says Argon agents are helping migrate C and C++ codebases to Rust, with some efforts reaching 800,000-plus lines, subject to auditing and testing. - A faster Rust video decoder, according to Google
In one example, Argon replaced 32,000 lines of SIMD code in a Rust port of libgav1; Google says the result ran 2.7 times faster with identical video output. - Memory savings at data-centre scale
Google says agents identified and applied fleet-wide memory optimisations that freed more than 300 TiB once rolled out, with estimated total savings of 500 TiB to 1 PiB. - A notable quantum-computing result
Google says Argon helped optimise a quantum algorithm subroutine, beating its published baseline by 40% in one example.
Why it matters
Argon’s significance is not just a bigger context window or a fresh benchmark score. Google describes using it for software migrations, infrastructure optimisation and vulnerability patching: jobs where the model’s output can affect real systems, and where review is not optional decoration.
The company says Argon can autonomously find, validate and patch critical software vulnerabilities, and that it is giving trusted defenders access to the model without cyber guardrails. That raises the stakes of the staged release. Google says it is strengthening safeguards against misuse and prompt injection, monitoring model reasoning and actions, and hardening sandboxed environments before broader access.
Our read
This is a substantial launch with unusually concrete examples of the work Google says its model can already tackle. The Rust-decoder example is especially legible: a claimed speed improvement, identical output and a specific engineering task, rather than the customary fog of “transformation”. Still, company demonstrations and benchmark scores are evidence of what Google reports, not a substitute for independent testing or production results.
For developers and organisations, the practical next step is to watch for wider access, actual API terms and independent evaluation. For defenders, the question is whether the phased programme produces useful security work without making powerful offensive capabilities too easy to obtain.
What to watch
- When paid API customers and Google AI Ultra subscribers receive access, and what the final pricing and limits are.
- Whether independent tests reproduce the coding, knowledge-work and cybersecurity results.
- What trusted cyber defenders report about vulnerability discovery, validation and patching in real deployments.
- How Google’s safeguards and monitoring perform as access expands.
Discussion spark: Should frontier models capable of finding and patching serious vulnerabilities be available to trusted defenders before the wider public, or does that approach create too much dependence on a company choosing who counts as trusted?
Sources and evidence
- Gemini 4 Argon: our next era of frontier intelligence (30 September 2026, 20:01 UTC)
not affiliated with, endorsed by, or operated by Google or Google DeepMind