Gemini 4 Argon benchmarks leak: tops knowledge work and long-context, mixed in coding
Overflame · reddit · 2026-10-14
An alleged official eval table for Gemini 4 Argon (methodology link points to deepmind.google) pits it against GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.5:
- Knowledge work: leads with Vals Index 68.9%, AutomationBench 51.3%, and Harvey's Legal Agent Benchmark 19.6% (rivals all under 7%).
- Agentic coding: tops DeepSWE v1.1 (77.9%) and Vibe Code Bench (91.9%), but trails on FrontierSWE v2 (55.0% vs GPT-6 Astra's 65.5%) and Terminal-bench 4.0 (Claude Opus 5.5: 66.4%).
- Long context: GraphWalks 256k–1M at 84.2%, well ahead of rivals (65–72%).
- Science/multimodal/security: first on LABBench 2 (88.8%), RiemannBench (76.0%), Chartography (71.6%), CWE-bench (68.0%); OSWorld-2.0 (69.2%) behind GPT-6 Astra (72.6%).
Overall Argon dominates agentic knowledge work, long context and multimodal understanding, while some coding and computer-use benchmarks remain contested.
Related event: Gemini 4 Argon Tops Multiple Benchmarks at Half the Cost(6 posts)→
More from Models
- Researcher using GLM gets Microsoft MSRC thanks for 23 Windows CVEs, joins Z.ai — Zai_org · 2026-10-14
- Mathematicians clash with OpenAI over how it handled Millennium Problem solutions — paulnovosad · 2026-10-14
- US labs are shipping SOTA nonstop while Chinese AI labs go unusually quiet — kimmonismus · 2026-10-14
- Liquid Claims 184 tok/s on My Mac — I Measured 174: The Speed Claim Holds — TheOyinbooke · 2026-10-14
- Dev Pays for Gemini Ultra, Walks Away Unimpressed by Gemini 4 Argon + Antigravity — yihui_indie · 2026-10-14
- NVIDIA Open-Sources a Chest CT Model Trained on 70,111 Scans That Writes Reports in Under 30 Seconds — MaziyarPanahi · 2026-10-14