ProductSep 28, 2026, 16:59 UTC

Leaked Gemini 4 Pro benchmarks top GPT-6 Astra and Claude Opus 5.5

What's the deal? A model listed as gemini-3.8-flash surfaced on LMArena in mid-September ‌2026, but community testers believe it is actually Gemini 4 Pro running under the internal codename Argon , an unannounced flagship. A leaked benchmark table circulating on X claims Gemini 4 Pro scores 88.7% on DeepSWE v1.1, 95.3% on Terminal-bench 2.1 and 86.8% on OSWorld-2.0, above OpenAI's GPT-6 Astra (74.1%, ~88% and 72.6% offline-partial, respectively,and Anthropic's Claude Opus 5.5 (74.2%, ~88% and 81.8% partial,. The same table lists a 2M-token context window and pricing of $2.25 / $11.25 per million tokens,versus $4 / $20 for Opus ‍5.5 and $10 / $50 for Astra. Community arena tests also show the suspected model beating Opus 5.5 by a large margin on 3D-physics prompts,such as the floatplane takeoff scene circulated in late September ‌2026.

Why it matters: Google has not confirmed the leak or a launch date. It lands days after Anthropic shipped Claude Opus 5.5 on September 22, ‌2026, and amid OpenAI's GPT-6 rollout,when the frontier race looked like a two-horse contest. If the numbers hold,Google DeepMind would re-enter the frontier as a capability-and-price leader,following the shelved Gemini 3.5 Pro and a roughly seven-month gap since Gemini 3.1 Pro. Earlier leak cycles have proved noisier than headlines suggested,so the results remain unverified until Google says otherwise.