Gemini 4 Argon tops coding benchmarks, adds 1M-token output

Google's frontier model beats Opus 5.5 and GPT-6 Astra on real-world SWE, but launch access is gated to Fairwind cyber testers for now.

Nowline Oct 1 2:00 AM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • New SOTA on real-world SWE

    Argon posts 77.9% on DeepSWE v1.1, edging Claude Opus 5.5 (74.2%) and GPT-6 Astra (74.1%). If it holds up on your repo, it's the new default to reach for on agentic coding and large refactors.

  • 1M-token output window

    Max output jumps from 64K to 1M tokens. You can generate an entire migration, a full test suite, or a book-length draft in one call instead of stitching chunks and re-priming context each time.

  • $2/$10 intro, then $4/$20

    Introductory pricing is $2 in / $10 out per million (cached input 95% off), matching Sonnet 5.5 and GPT-6.1 Sol. But it doubles to $4/$20 once the intro window closes, so budget for the step-up before you wire it into anything hot.

  • Cyber defenders get it first

    Argon ties for #1 on CWE-bench v1 (68%) for autonomous vulnerability remediation, so Google is shipping it to trusted cyber defenders via the Fairwind Program before anyone else. You can't call it today unless you're in that program; broad API and Google AI Ultra access is 'rolling out soon' with no date, so watch AI Studio.

  • Autonomous Rust migrations, proven at scale

    Google ran Argon agents to migrate 800K+ lines of C/C++ to Rust (including Fuchsia OS), rewrite 32K lines of libgav1 SIMD for a 2.7x speedup, and free 300+ TiB of fleet memory. The weekend unlock once it opens up: point it at legacy code and let it port and optimize.