Gemini 4 Pro (Leaked Benchmark)
Repository-scale 2M-token refactoring (88.7% DeepSWE v1.1), headless shell execution (95.3% Terminal-Bench 2.1), and multimodal desktop agent autonomy (86.8% OSWorld 2.0).
SWE-bench Verified
88.7%
MMLU-Pro
93.8%
MATH-500
97.8%
Context Window
2,000,000 tokens
| Specification | Value |
|---|---|
| Developer / Lab | Google DeepMind |
| Parameters Total | Proprietary Omni-Modal Unified Diffusion Transformer |
| Parameters Active | Dynamic Speculative MoE Backbone on TPU v6 Trillium |
| Architecture Type | Unified Multi-Token Diffusion Transformer + Recursive Self-Improvement (RSI) + RingAttention 2M Context |
| License | Proprietary Commercial API (Google AI Studio, Vertex AI) |
| Input Pricing (per 1M) | $2.25 |
| Output Pricing (per 1M) | $11.25 |
Head-to-Head Comparisons with Gemini 4 Pro (Leaked Benchmark)
Gemini 4 Pro (Leaked Benchmark) vs DeepSeek-V4.1-Flash-1017 →Gemini 4 Pro (Leaked Benchmark) vs Perplexity Decider v1.2 →Gemini 4 Pro (Leaked Benchmark) vs NVIDIA Nemotron Labs Dory 1.0 →Gemini 4 Pro (Leaked Benchmark) vs Gemini 4 Carbon →Gemini 4 Pro (Leaked Benchmark) vs Nano Banana 3 Pro →Gemini 4 Pro (Leaked Benchmark) vs Claude Opus 5.5 (Canary Checkpoint) →