FlowtandBriefAI · Software · ResearchMeet Flowtand

Google announces Gemini 4 Argon, a long-horizon model with a 1M output token limit

Announcement

Google DeepMind

Browse Models topic page →

Google announced Gemini 4 Argon, a frontier model it describes as built for long, multi-step work in software engineering, enterprise knowledge tasks and cybersecurity defense. It is rolling out first to trusted cyber defenders via the Fairwind Program, with the 1M output token limit, introductory API pricing and vendor-reported benchmark scores. Google says broader access will follow phased testing and guardrail work.

What Google announced

Google announced Gemini 4 Argon as a frontier model aimed at long-horizon, multi-step work across software engineering, legal and finance knowledge work, and cybersecurity defense. The company says the model's output token limit rises from 64K to 1M, which it frames as headroom for deeper reasoning within a single trajectory. Access is phased: Argon is rolling out to trusted cyber defenders through the Fairwind Program, and Google says it is engaging with a U.S. government pre-release access process. Google reports state-of-the-art scores on DeepSWE v1.1 (77.9%), LVBench (91.7%) and AutomationBench (51.3%), plus internal coding, quantum and data-center memory examples. It states launch pricing of $2 per million input tokens and $10 per million output tokens, with a later $4/$20 rate. Every performance figure is vendor-reported.

A scenario for a security engineering lead

A security engineering lead could consider requesting Fairwind access to test whether Argon helps with defensive triage: having the model inspect a curated set of known vulnerabilities in a sandboxed copy of internal code, propose patches and produce evidence that each patch addresses the issue. The plausible premise is Google's claim that Argon can find, validate and patch vulnerabilities, and that Wiz is using it through Scan for Good. This would be a controlled internal evaluation, not an endorsement of the model's real-world effectiveness.

How to trial and measure it

A team could run a small bake-off against its current model or a manual baseline on a fixed set of tasks, such as patch proposals for a known vulnerability list, with identical prompts and reviewers blind to which system produced each result. Record acceptance rate, time to a usable patch, reviewer edits and false positives, then compare against the baseline. Include failure checks: verify patches compile and pass tests, and require human review before any change reaches production. Treat Google's benchmark scores as context only, not as a substitute for your own measurements.

Trade-offs and unknowns

Longer outputs can mean longer runtimes and higher cost per task, so the $2/$10 introductory pricing needs to be modeled against expected token volume before committing. The broad release date, regional availability and final terms are not settled in the source. Google says trusted defenders get Argon without cyber guardrails while general release will be more guarded, so internal use cases may not match public behavior. A team should not use Argon for autonomous production changes, unreviewed legal or financial outputs, or any workflow lacking human oversight until independent evidence exists.

Context and limitations

  • Based on a short excerpt of a vendor announcement; independent testing and full benchmark methodology were not reviewed.
  • Performance figures, savings and internal case studies are Google's own claims, not independently verified.
  • General developer, enterprise and consumer availability, regional terms and final pricing are not established here.
  • Based on bounded source text, not independent testing. Practical suggestions are editorial proposals.