Gemini 4 Hydrogen vs Claude Sonnet 5.5: Did Google Beat Claude?
Google has officially announced Gemini 4 Hydrogen, its latest fast model, which competes with Claude Sonnet 5.5. The benchmarks Google shared show Gemini 4 Hydrogen beating Claude Sonnet 5.5 on most tests, but in overall intelligence Claude Sonnet 5.5 is slightly ahead. Where Gemini 4 Hydrogen stands out is efficiency. Here is how the two models compare.
Google's benchmarks | |||
| Benchmark |  Gemini 4 Hydrogen |  Claude Sonnet 5.5 | Difference |
|---|---|---|---|
| SWE-sweep | 13.5% | 3.1% |  +10.4 pts |
| Terminal-Bench 4.1, cheating-adjusted | 57.0% | 63.9% |  +6.9 pts |
| Agents' Last Exam v1.1 | 71.6% | 60.4% | +11.2 pts |
| AutomationBench-AA | 76.1% | 71.8% | +4.3 pts |
| Vending-Bench 2, bank balance | $14,549 | $11,165 | +$3,384 |
| OSWorld 2.1, offline | 77.6% | 83.9% | +6.3 pts |
| MKA, Climb WSR | 53.4% | 52.0% | +1.4 pts |
| Chartography | 66.6% | 61.6% | +5.0 pts |
| Terminal-Bench Science 0.1 | 51.4% | 59.9% | +8.5 pts |
| HLE-Diamond | 58.8% | 53.3% | +5.5 pts |
Artificial Analysis benchmarks | |||
| Intelligence Index v5.0.1 | 54 at xhigh | 55 at max | +1 |
| AA-Briefcase v1.1 | 1567 Elo at xhigh | 1816 Elo at max | +249 Elo |
Cost and speed at top settings | |||
| Cost per Intelligence Index task | $3.16 | $10.02 | 68% cheaper |
| Output tokens across the Index | 304M | 335M | 9% fewer |
| Output speed | 190 tokens/s | 139 tokens/s | 37% faster |
| Time to first answer token | 37.8 seconds | 240.2 seconds | 84% sooner |
| Time per Index task | 13.5 minutes | 34.1 minutes | 60% less |
Intelligence Index by effort | |||
| Max | Not offered | 55 | |
| xhigh | 54 | 50 | +4 |
| High | 51 | 42 | +9 |
| Medium | 41 | Not rated | |
| Low | 27 | 26 | +1 |
Cost per task by effort | |||
| Max | Not offered | $10.02 | |
| xhigh | $3.16 | $4.09 | 23% cheaper |
| High | $1.87 | $1.64 | 12% cheaper |
| Medium | $0.80 | Not rated | |
| Low | $0.29 | $0.51 | 43% cheaper |
AA-Briefcase by effort | |||
| Max | Not offered | 1816 Elo | |
| xhigh | 1567 Elo | 1742 Elo | +175 |
| High | 1497 Elo | 1634 Elo | +137 |
| Medium | 1296 Elo | Not rated | |
| Low | 947 Elo | 1268 Elo | +321 |
Pricing and specs | |||
| Input, per 1M tokens | $1.50 | $2 | 25% cheaper |
| Output, per 1M tokens | $7.50 | $10 | 25% cheaper |
| Cached input, per 1M tokens | $0.075 | $0.20 | 63% cheaper |
| Context window | 1M tokens (input + output) | 1M tokens | |
| Max output | 64K tokens, up to 1M with long decode | 128K tokens | 2x by default |
| Reasoning levels | Low to xhigh | Low to max | |
| Default effort | High | High | |
| Inputs | Text, images, audio, video | Text, images, PDFs | |
| Knowledge cutoff | March 2026 | June 2026 | 3 months newer |
| Released | November 12, 2026 | September 28, 2026 | 45 days newer |