Leaked benchmark table shows Gemini 4 Pro ahead of Opus 5.5 and Astra
A leaked table lists Gemini 4 Pro at DeepSWE 88.7, Terminal-Bench 95.3, OSWorld 86.8, 2M context and $2.25/$11.25 pricing, all above Opus 5.5 and GPT-6 Astra, with the author noting it is unverified.
GEMINI 4 PRO LEAKS ARE WILD. 👀 LEAKED Gemini 4 Pro vs current flagships: DeepSWE v1.1 88.7 — G4 Pro (LEAK) 74.2 — Opus 5.5 74.1 — GPT-6 Astra Terminal-Bench 2.1 95.3 — G4 Pro (LEAK) 87.64 — Opus 5.5 87.27 — GPT-6 Astra OSWorld 2.0 86.8 — G4 Pro (LEAK) 81.8 — Opus 5.5 (partial) 72.6 — GPT-6 Astra (offline) GDPval-AA v2.1 2064 — G4 Pro (LEAK) 1846 — Opus 5.5 1542 — GPT-6 Astra Context 2M — G4 Pro (LEAK) 1.1M — GPT-6 Astra 1M — Opus 5.5 Price / 1M tokens $2.25 / $11.25 — G4 Pro (LEAK) $4 / $20 — Opus 5.5 $10 / $50 — GPT-6 Astra And this is the part that makes me pause… If these Gemini 4 numbers are real, what exactly are we witnessing? Google has Gemini 4 in post-training, while DeepMind has increasingly talked about AI-assisted model development and recursive evaluation/improvement. Is that connected to the delays? RSI? More aggressive post-training? AI-driven evaluation loops? Or simply a massive amount of compute? We don't know. The Gemini 4 numbers are LEAKED and unverified. The RSI connection is speculation, not confirmed. But if independent testing validates even most of this… the next frontier-model race could look VERY different.🔥