SYLIN

Gemini Argon

by Google · Model · Language & code models

mixed

Strong at launch and low on hallucinations, but retired and overtaken by newer Gemini 4 models.

Took #1 on Arena text

AccessAPI

Why we say so

Artificial Analysis found the lowest hallucination rate (about 15%), Vals Index ranked it first, and Blueprint-Bench rated it best at floor plans. Its Artificial Analysis score was a few points behind Opus 5.5, though, and Carbon scored 40.0% versus Argon's 26.3% on FrontierCode 2.0, a figure from Cognition. HN users call it amazing but very slow in Antigravity and say it runs long on simple questions. Google has since retired it in favor of Neon, which one commenter finds as good but faster.

Ranks #3 for chat · #3 for knowledge work

Evidence

  1. Nov 19
    Same price as Carbon, doubling Dec 1; Google says it gives Gemini 4 intelligence with faster responses and fewer tokens per turn
  2. Nov 11
    #2 on Arena text (1532)
  3. Oct 18
    Google releases Gemini 4 Argon to paid Pro/Ultra plans and paid API; free plan excluded
  4. Oct 18
    Google's benchmarks show it beating GPT-6 Astra and Claude Opus 5.5 on various knowledge-work, agentic-coding and science tests
  5. Oct 18
    Can output 1 million tokens in one response, roughly 10x other frontier models' 64K-128K output limit
Show all 13
  1. Oct 18
    Best model on Blueprint-Bench, converting apartment photos to floor plans, nearly matching the human baseline
  2. Oct 18
    Ranked first on Vals Index and much cheaper than second and third place
  3. Oct 18
    Ranked first on a medical reasoning benchmark with clinical images
  4. Oct 4
    Google's new frontier model, on average beating GPT-6 Astra and top Claude models on agentic coding, science, math and computer use
  5. Oct 4
    Ranked first on long-horizon agentic coding, Vals index, financial research and legal work benchmarks
  6. Oct 4
    Tied for second with GPT-6 Astra and Claude Fable on Artificial Analysis index, a few points behind Opus 5.5, but cheaper
  7. Oct 4
    Only for trusted cyber defenders for now; later paid API customers and Google AI Ultra subscribers first
  8. Oct 4
    Lowest hallucination rate on AA-Omniscience, about 15% versus about 50% for GPT-6 and 60-70% for Claude

Alternatives

Gemini CarbonArgon preferred for most coding tasks; Carbon preferred for architecture and deep root-cause analysis of complex issues
Gemini NeonCommenter finds it as good as Gemini Argon but faster and (they think) cheaper
Gemini HydrogenRoboflow vision bench at high effort: 29.3s per image vs 51.4s Argon and 132.3s Carbon, though Carbon detects more accurately
DeepSeek FlashRanks third on KernelBench CUDA, below the latest Claude models but ahead of Gemini 4 Argon, Claude Fable and GPT-6 Astra
Built on Gemini Argon