Gemini 4 Carbon vs Claude Opus 5.5: Google Carbon Beats Opus
With its latest model, it looks like Google has finally beaten Claude Opus 5.5, the best frontier model on the market right now. Here is how the two compare in benchmarks. Just remember that models that do well in benchmarks don't always do as well in real-world use. With that in mind, let's see how they compare.
Google benchmarks
| Benchmark | Gemini 4 Carbon | Claude Opus 5.5 | Difference |
|---|---|---|---|
| AA Intelligence Index v5.0 | 59 | 58 | Carbon +1 |
| Vals Index v2.2 | 70.4% | 70.2% | Carbon +0.2 points |
| SWE-sweep | 9.6% | 3.6% | Carbon +6.0 points |
| FrontierSWE 2 | 61.2% | 62.3% | Opus +1.1 points |
| Terminal-Bench 4.0 (AA) | 65.2% | 59.6% | Carbon +5.6 points |
| Agents' Last Exam v1.1 | 77.4% | 71.6% | Carbon +5.8 points |
| OSWorld-2.1 (offline subset) | 81.5% | 87.2% | Opus +5.7 points |
| AutomationBench (Zapier) | 53.9% | 42.5% | Carbon +11.4 points |
| Vending-Bench 2 (bank balance) | $16,484 | $9,235 | Carbon +$7,249 |
| Chartography | 72.5% | 66.3% | Carbon +6.2 points |
| Blueprint-Bench 2 | 58.6% | 51.2% | Carbon +7.4 points |
| GraphWalks 256K to 1M | 96.2% | 66.8% | Carbon +29.4 points |
| RiemannBench | 80.0% | 69.6% | Carbon +10.4 points |
| Terminal-Bench Science 0.1 (AA) | 64.8% | 59.0% | Carbon +5.8 points |
| CWE-bench v1 | 71.0% | 67.0% | Carbon +4.0 points |
| ExploitGym | 54.8% | 34.5% | Carbon +20.3 points |
| SEC-bench Pro | 88.2% | - | Not reported for Opus |
Artificial Analysis benchmarks
| Benchmark | Gemini 4 Carbon (High) | Claude Opus 5.5 (Max) | Difference |
|---|---|---|---|
| Intelligence Index v5.0 | 59 | 58 | Carbon +1 |
| AA-Briefcase v1.1 | 1609 Elo | 1800 Elo | Opus +191 Elo |
| AutomationBench-AA | 78% | 70% | Carbon +8 points |
| Terminal-Bench 4.0 | 65% | 60% | Carbon +5 points |
| FrontierCode 2.0 | 41% | 46% | Opus +5 points |
| AA-Omniscience v1.1 | 54 | 48 | Carbon +6 |
| GDP.pdf | 28% | 26% | Carbon +2 points |
| HLE-Diamond | 69% | 68% | Carbon +1 point |
| CritPt v2 (Hard Subset) | 63% | 61% | Carbon +2 points |
| Terminal-Bench Science 0.1 | 65% | 59% | Carbon +6 points |
Cost and speed
| Metric | Gemini 4 Carbon (High) | Claude Opus 5.5 (Max) | Difference |
|---|---|---|---|
| Cost per Intelligence Index task | $7.76 | $10.04 | Carbon about 23% cheaper |
| Cost to run the Intelligence Index | $8,240 | $8,910 | Carbon about 8% cheaper |
| Output tokens per task | 123K | 173K | Carbon about 29% fewer |
| Reasoning tokens per task | 72K | 120K | Carbon 40% fewer |
| Output tokens to run the Index | 205M | 218M | Carbon about 6% fewer |
| Output speed | 92 tokens/s | 98 tokens/s | Opus about 6% faster |
| Time to first token | 100.31 s | 663.66 s | Carbon starts about 6.6x sooner |
| Time per task | 1342.93 s | 1147.38 s | Opus finishes about 15% sooner |
Pricing and specs
| Specification | Gemini 4 Carbon | Claude Opus 5.5 | Difference |
|---|---|---|---|
| Input, per 1M tokens | $2Introductory, then $4 | $4 | Carbon half the price for now |
| Cached input, per 1M tokens | $0.10Introductory, then $0.20 | $0.20 | Carbon half the price for now |
| Output, per 1M tokens | $10Introductory, then $20 | $20 | Carbon half the price for now |
| Context window | 1M tokens | 1M tokens | Same |
| Max output | Not published | 128K tokens300K with the Batch API | - |
| Reasoning | Yes | Adaptive, always on | Different controls |
| Input | Text, image, speech, video | Text, image | Carbon takes audio and video |
| Public API | Not yetCyber defenders only | Yesclaude-opus-5-5 | Opus |
| Released | October 28, 2026 | September 22, 2026 | Carbon 36 days newer |
The catch with the price
Carbon's $2 and $10 is an introductory price. After it ends, Carbon costs $4 and $20 per million tokens, exactly the same as Opus 5.5, and Google hasn't said when that happens. On top of that, you can't use Carbon yet unless you are one of the cyber defenders Google picked for early access. Opus 5.5 is available today on the Claude API, Amazon Bedrock, Google Cloud and Microsoft Foundry.