Head-to-head · Updated Nov 3, 2026
Gemini 4 Carbon vs Claude Opus 5.5
Google’s Gemini 4 Carbon and Anthropic’s Claude Opus 5.5 on independent tests for professional work, cost and speed.
Short answer
Gemini 4 Carbon leads on 6 of 10 quality scores, including the overall Intelligence Index.
- FinanceGemini 4 Carbon62.6 vs 59.4
- Legal workGemini 4 Carbon62.0 vs 58.0
- EngineeringGemini 4 Carbon64.7 vs 64.6
- Strategy & operationsClaude Opus 5.561.4 vs 60.9
- Coding agentsGemini 4 Carbon65.2% vs 59.6%
- Fewer made-up answersGemini 4 Carbon11.4% vs 50.3%
- Lower costGemini 4 Carbon$8,240 vs $8,910
- SpeedClaude Opus 5.596 tok/s vs 94 tok/s
Scores, cost and speed
| Measure | Gemini 4 Carbon (High)Limited early access | Claude Opus 5.5 (Max, Default Fallback) |
|---|---|---|
| Overall | ||
| Intelligence IndexCombined score across nine evaluations | Leads: 59.4 | 58.4 |
| Professional work | ||
| Finance & accounting | Leads: 62.6 | 59.4 |
| Legal | Leads: 62.0 | 58.0 |
| Healthcare & medical | — | 57.9 |
| Engineering | Leads: 64.7 | 64.6 |
| Strategy & operations | 60.9 | Leads: 61.4 |
| Economics | 61.0 | Leads: 62.3 |
| Specific tests | ||
| Terminal-Bench 4.0Agentic coding tasks | Leads: 65.2% | 59.6% |
| FrontierCode 2.0Hard programming problems | 40.6% | Leads: 46.4% |
| Humanity’s Last ExamExpert-level questions | — | 61.4% |
| HealthBenchHealth conversations graded by physicians’ rubrics | — | 57.0% |
| GDPval-AAKnowledge-work deliverables, Elo | 1733 | Leads: 1866 |
| Hallucination rateAA-Omniscience; lower is better | Leads: 11.4% | 50.3% |
| Cost & speed | ||
| Input priceUS$ per 1M tokens | $4 | $4 |
| Output priceUS$ per 1M tokens | $20 | $20 |
| Cost to run the indexUS$, includes reasoning tokens | Leads: $8,240 | $8,910 |
| Output speedMedian tokens per second | 94 tok/s | Leads: 96 tok/s |
| Context windowTokens | 1M | 1M |
Scores, prices and speeds: Artificial Analysis, independently measured, checked Nov 3, 2026. Model names include the tested reasoning setting. Stubbi has not rerun these evaluations. Consumer apps such as ChatGPT, Claude and Gemini may use other models and add features like search and memory.
Test them on your own work
Public scores can’t show how each model handles your documents, clients or standards. Tell us the task, and we’ll consider a side-by-side comparison with the outputs you can judge.