Gemini 4 Neon benchmark leak claims sweep over Claude Opus 5.5 and GPT-6.2 Sol
kristoph · x · 2026-11-19
Developer kristoph notes Google went from "written off 6 months ago" to "winning hard," sharing a benchmark table comparing Gemini 4 Neon against Claude Opus 5.5 and GPT-6.2 Sol (attributed to a DeepMind evals page; model names unverified).
Key claimed results, with Gemini 4 Neon leading most categories:
- Capability indices: AA Index 61 (vs 59/57), Vals Index 70.9%
- Agentic coding: CursorBench 65.4%, Terminal-bench 65.2%, SWE-sweep provider harness 39.0% (vs 10.1%/8.3%)
- Knowledge work: AutomationBench 56.3%, Agents' Last Exam 75.8%
- Science & math: RiemannBench 84.0%, HLE-Diamond 71.3%
- Long context: GraphWalks (256k–1M) 95.2% vs 66.8%/80.1%
- Cybersecurity: SEC-bench Pro 90.5%, WebExploitBench 51.3% (vs 37.1%)
- Claude Opus 5.5 narrowly wins Terminal-bench and FrontierCode 2.0; GPT-6.2 Sol leads Chartography.
More from Models
- Gemini 4 Neon costs $6.44 per task, 17% below Carbon but 1.6x GPT-6.2 Astra — ArtificialAnlys · 2026-11-19
- Gemini 4 Neon hits 59 on Coding Agent Index, matching Claude flagships at half the cost — ArtificialAnlys · 2026-11-19
- Gemini 4 Neon sets record 1771 Elo on Presentation, still trails Claude on Briefcase — ArtificialAnlys · 2026-11-19
- Gemini 4 Neon matches Claude Opus 5.5 on coding agents at roughly half the cost — ArtificialAnlys · 2026-11-19
- Reddit: Gemini 4 Neon beats Claude Opus 5.5 on most benchmarks — Endonium · 2026-11-19
- Reddit buzz: Gemini 4 Neon reportedly beats Opus 5.5 at lower cost — Electronic-Weekend66 · 2026-11-19