Gemini 3 Deep Think
Last updated: April 22, 2026
Specs
| Vendor | Google DeepMind |
|---|
| Released | January 28, 2026 |
|---|
| Context window | 1M tokens |
|---|
| Input | $14 / M tokens |
|---|
| Output | $56 / M tokens |
|---|
Benchmarks
| mmlu | 92.1 |
|---|
| gpqa | 84.3 |
|---|
| humanEval | 92.4 |
|---|
| arenaElo | 1429 |
|---|
| sweBench | 64 |
|---|
| sweBenchPro | 55 |
|---|
| liveCodeBench | 80 |
|---|
| aiderPolyglot | 75 |
|---|
| terminalBench | 45 |
|---|
| mmluPro | 87 |
|---|
| hle | 30 |
|---|
| aime | 95 |
|---|
| math500 | 98.5 |
|---|
| mmmu | 80 |
|---|
Strengths
- Reasoning-tuned Gemini: close to o4 on GPQA
- 1M context keeps long problems tractable
- Strong performance on competition math (IMO-style)
Weaknesses
- Slower than vanilla Gemini 3 Pro
- Availability behind Google AI Ultra subscription
- Higher price than Pro with smaller quality gap on everyday tasks
Best for
- Scientific reasoning under Google Cloud
- Competitive math and olympiad training
- Long-context research assistants
Compare with other AI models