Gemini 3 Pro Preview
Preview Gemini flagship for complex reasoning, coding, and rich multimodal prompts
Shogo Agent Index
Supported#48 of 9244.7range 37.9–51.6 · 58% of weight measured
Agentic coding
48
57% measured
Long-horizon autonomy
60
80% measured
Knowledge-work agents
34
68% measured
Computer use
33
estimated
Security & context learning
43
50% measured
Price
- Input
- $2.00/M
- Output
- $12.00/M
- Cache read
- —/M
- 3:1 blend
- $4.50/M
- Agent-effective
- $2.48/M
Agent-effective = 20 input tokens per output token, 85% cache hits. Why.
Agentic benchmarks
| Benchmark | Score | Normalised | Index weight | Harness / source |
|---|---|---|---|---|
| SWE-bench Verified | 72.9% | 8% | source | |
| Terminal-Bench 2.0 | 69.4% | 8% | Ante | |
| DeepSWE | est. | 6% | predicted 44 | |
| FrontierSWE | est. | 5% | predicted 12 | |
| CursorBench | est. | 4% | predicted 35 | |
| GSO-Bench | 18.6% | 4% | OpenHands | |
| METR Time Horizons | 71.0% | 10% | source | |
| Vending-Bench 2 | $5,478 | 6% | source | |
| PostTrainBench | est. | 4% | predicted 23 | |
| APEX-Agents | est. | 8% | predicted 41 | |
| GDPval | 40.3% | 7% | source | |
| Remote Labor Index | 1.3% | 5% | source | |
| DeepResearch Bench | 46.3% | 5% | source | |
| OSWorld 2.0 | est. | 6% | predicted 7 | |
| OSWorld | est. | 4% | predicted 71 | |
| Cybench | est. | 5% | predicted 70 | |
| CL-bench | 15.8% | 5% | source |
Other benchmarks
| Benchmark | Category | Score | Source |
|---|---|---|---|
| AlgoTune | Coding | 1.83× | source |
| ARC-AGI | Reasoning | 75.0% | source |
| ARC-AGI-2 | Reasoning | 31.1% | source |
| BALROG | Long-horizon | 58.1% | source |
| CritPt | Reasoning | 6.9% | source |
| GeoBench | Multimodal | 84.0% | source |
| GPQA Diamond | Reasoning | 92.6% | source |
| Humanity's Last Exam | Reasoning | 37.5% | source |
| OTIS Mock AIME | Math | 91.4% | source |
| ProofBench | Math | 20.0% | source |
| SimpleBench | Reasoning | 76.4% | source |
| VPCT | Multimodal | 91.0% | source |
| WebDev Arena | Coding | 1439 | source |