Model cards

gemini-4-argon

Vendor google Tier - Price per million tokens, in / out $2 / $10 Released 2026-09-30 Context 1,048,576 Max output 1,000,000 Vendor page

Last updated 2026-10-26 How this is made

Meant to excel at

• Google's flagship reasoning model for complex software engineering, advanced math and science, and autonomous multi-step agents.

  Vendor's own site ai.google.dev

• Long decode: outputs can reach about 1M tokens, versus a 64K default; useful for very long generations and long-running agents.

  Vendor's own site ai.google.dev

• Very low hallucination rate (15% on AA-Omniscience); it declines to answer instead of guessing.

  Independent evaluator eesel.ai

• Leads finance, legal and business agent benchmarks and long-document retrieval (GraphWalks 256K to 1M), mostly third-party leaderboards.

  Independent evaluator eesel.ai

• Ranked first on AutomationBench-AA multi-step automation (77.5%), ahead of Claude Sonnet 5.5 and Opus 5.5.

  Independent evaluator eesel.ai

• Google used its agents to translate C/C++ to Rust, including over 800K lines of Fuchsia kernel code.

  Vendor claim, relayed deeplearning.ai

• Debugged Vulkan code by dumping framebuffers to PNG and using false-color shaders, per one HN user.

  Practitioner report news.ycombinator.com

• Ranked first on the Vals Index (68.90% as of Oct 12, ahead of Sonnet 5.5 at 67.04%) and on Vals Finance Agent v2 (65.4%).

  Independent evaluator letsdatascience.com

Watch out for

• Trails Anthropic's newest models on terminal and agentic coding (Terminal-Bench 4.0 57.1%, fourth); index five points behind Opus 5.5.

  Independent evaluator eesel.ai

• Token-hungry: about 62K output tokens per task versus 27K for GPT-6 Astra; list price rises from $2/$10 to $4/$20.

  Independent evaluator eesel.ai

• At high effort it can spend the whole 65,536-token default output budget on reasoning and never write the answer file.

  Practitioner report datacamp.com

• Access restricted: vetted Fairwind cyber defenders first, then paid API and Ultra; no Live API, no explicit caching.

  Vendor claim, relayed deeplearning.ai

• Thinks a lot: about 24k output tokens and 4 minutes per small fix vs 3k and 32s for Sonnet 5.5; one fix timed out; pulls many unrelated files into context.

  Practitioner report news.ycombinator.com

• Independent rankings split: first on Vals professional workflows but only tied eighth of 216 on the AA index (53 vs 58 for Opus 5.5).

  Independent evaluator letsdatascience.com

Benchmarks

Coding percentile 0.68

Agentic coding percentile 0.88

Overall percentile 0.85

Percentile is the share of tracked models this model beats on that benchmark, averaged across the group. Only benchmarks at least four tracked models share are scored.

Practitioner reports

• +/-

  Plotly's 43-question analytics test had Argon at 41/43 and all five hidden-data issues found, using about 46 queries per question; tied with Opus 5.5. Single run, one harness.

  Task: data-analysis | Compared with claude-opus-5.5 | 2026-10-17

  Practitioner report plotly.com
• -

  A Pro subscriber found Argon generates code much more slowly than Gemini 3.1 Pro did and threatened to leave Google's ecosystem.

  Compared with gemini-3.1-pro | 2026-10-15

  Practitioner report discuss.ai.google.dev
• +

  An educator had Argon audit an app he had built with Opus 5.5 and it found several bugs, one that could have exposed his webhook, which he patched. He notes the real risk was low and he still prefers other models for some coding.

  Task: code-review | 2026-10-15

  Practitioner report favtutor.com
• -

  An Ultra subscriber says Argon burns through quota quickly and wants Gemini 3.1 Pro back.

  Compared with gemini-3.1-pro | 2026-10-14

  Practitioner report discuss.ai.google.dev
• +/-

  One HN user's personal eval of 11 small bug fixes: Argon solved 10 of 11, at a median 4 minutes and $0.52 per run vs $0.18 and 32s for Sonnet 5.5 and $0.09 and 1.5 min for Sol. Costs came from a tool that undercounted Argon pricing; one commenter said Opus 5.5 starts editing sooner.

  Task: bug-fix | Compared with claude-sonnet-5.5 | 2026-10-14

  Practitioner report news.ycombinator.com
• +/-

  Canvas built a working booking page with filters and validation, and Argon handled a data-decline analysis well and refused to claim causation; on document reconciliation it labeled all eight claims correctly but one summary overstated what the packet said.

  2026-10-14

  Practitioner report analyticsvidhya.com
• -

  Reviewer notes Argon wins most non-coding benchmarks but is last among cited frontier models on Terminal-bench 4.0 and trails on FrontierSWE v2; advises keeping coding pipelines on Claude or GPT rather than migrating.

  Task: feature | Compared with claude-opus-5.5 | 2026-10-02

  Practitioner report dev.to
• +/-

  Artificial Analysis data cited by the same article shows Argon uses about 62k output tokens per task vs 27k for GPT-6 Astra, so its low intro price does not make it cheaper per task at regular pricing.

  Task: agentic-long-running | Compared with gpt-6-astra | 2026-10-02

  Practitioner report dev.to
• +/-

  Google's showcase work for Argon is code comprehension and migration (kernel comprehension, a SIMD rewrite), not greenfield development; commentators warn labs can optimize for leaderboards.

  Task: refactor | 2026-10-02

  Practitioner report dev.to
• +

  An early-access user says it is the first Gemini they can hand complex research tasks to, though not fully reliable. (HN comment relayed by eesel.)

  undated

  Practitioner report eesel.ai
• -

  In a single-file robot-arm animation test it handled the planned detour and honestly flagged an unreachable target, but later clipped through obstacles; Opus 5.5 kept the constraints better.

  Compared with claude-opus-5.5 undated

  Practitioner report datacamp.com
• -

  At high reasoning effort the run exhausted the default output budget on thinking and produced no file; the author had to rerun at medium effort.

  Compared with claude-opus-5.5 undated

  Practitioner report datacamp.com
• -

  Anonymous insiders reportedly told Bloomberg it benchmarks well but gets stuck on real business codebases, especially frontend work; Google called this inaccurate.

  undated

  Practitioner report tech.ifeng.com