Gemini 4 Carbon vs Gemini 4 Hydrogen
| Gemini 4 Carbon | Gemini 4 Hydrogen | |
|---|---|---|
| Announced | ||
| Released | ||
| Input / 1M tokens | $2$4 from 1 Dec 2026 | $1.50 |
| Cached input / 1M tokens | $0.10$0.20 from 1 Dec 2026 | $0.075 |
| Output / 1M tokens | $10$20 from 1 Dec 2026 | $7.50 |
| Context window | 1M | 1M |
| Output limit | 65K | 65K |
| Benchmarks | ||
| Terminal-Bench Science 0.1 | 64.8%1 | 51.4%2 |
| OSWorld 2.1 Offline set (82 tasks), partial score | 81.5%3 | 77.6%4 |
| Chartography No tools | 72.5%5 | 66.6%6 |
| Agents' Last Exam v1.1 Score, all tasks (122) | 77.4%7 | 71.6%8 |
| Gemini 4 Carbon | Gemini 4 Hydrogen | |
|---|---|---|
| Other benchmarks | ||
| HLE-Diamond No tools | No data | 58.8%9 |
| Terminal-Bench 4.0 | 65.2%10 | No data |
| AutomationBench 1.0.6 Private set | 53.9%11 | No data |
| AutomationBench 1.0.6 AutomationBench-AA | No data | 76.1%12 |
| FrontierSWE v2 | 61.2%13 | No data |
| SEC-bench Pro 260505 | 88.2%14 | No data |
| GraphWalks BFS, 256K + 1M problems | 96.2% | No data |
Notes
- Terminal-Bench Science 0.1, Gemini 4 Carbon: Run by Artificial Analysis · mini-swe-agent harness · high thinking · average of 3 runs
- Terminal-Bench Science 0.1, Gemini 4 Hydrogen: mini-swe-agent vision harness · highest thinking · average of 5 attempts per task
- OSWorld 2.1, Offline set (82 tasks), partial score, Gemini 4 Carbon: Gemini CUA harness · highest thinking · max of 3 runs, one attempt per run
- OSWorld 2.1, Offline set (82 tasks), partial score, Gemini 4 Hydrogen: Gemini CUA harness · highest thinking · max of 3 runs, one attempt per run
- Chartography, No tools, Gemini 4 Carbon: Run by Surge AI · Surge AI's public leaderboard
- Chartography, No tools, Gemini 4 Hydrogen: highest thinking
- Agents' Last Exam v1.1, Score, all tasks (122), Gemini 4 Carbon: ALE-Claw harness · 5-hour window · October 17, 2026 task files · safety filters on
- Agents' Last Exam v1.1, Score, all tasks (122), Gemini 4 Hydrogen: Antigravity-CLI (agy) version 1.4.8 harness · highest thinking · 5-hour window · October 17, 2026 task files · safety filters on
- HLE-Diamond, No tools, Gemini 4 Hydrogen: Run by Artificial Analysis · highest thinking
- Terminal-Bench 4.0, Gemini 4 Carbon: Run by Artificial Analysis · mini-swe-agent harness · high thinking · average of 3 runs
- AutomationBench 1.0.6, Private set, Gemini 4 Carbon: Run by Zapier · set named in the evaluation methodology the post cites
- AutomationBench 1.0.6, AutomationBench-AA, Gemini 4 Hydrogen: Run by Artificial Analysis · highest thinking
- FrontierSWE v2, Gemini 4 Carbon: Run by Proximal · Proximal's public leaderboard
- SEC-bench Pro 260505, Gemini 4 Carbon: Antigravity harness · pass@1, mean of 3 attempts · 183-task JavaScript engine split (May 2026) · LLM judge for crashes in both builds