Across 11 shared benchmarks, Claude Opus 4 leads overall: GPT-5 wins 5, Claude Opus 4 wins 6, with 0 ties and an average score difference of +3.45.
OpenAI · 2025-08-07 · Foundation model
Anthropic · 2025-05-23 · Reasoning model
Grouped by capability, sorted by largest gap within each. 11 shared benchmarks.
| Benchmark | GPT-5 | Claude Opus 4 | Diff |
|---|---|---|---|
| IMO-ProofBench | 592 / 16thinking | 2.9016 / 16thinking | +56.10 |
| FrontierMath | 26.3014 / 60Thinking High (With Tools) | 4.5039 / 60 | +21.80 |
| AIME2025 | 61.9080 / 106 | 75.5065 / 106 | -13.60 |
| FrontierMath - Tier 4 | 12.5029 / 80Thinking High (No Tools) | 072 / 80Normal (No Tools) | +12.50 |
| Simple Bench | 56.708 / 27high | 58.807 / 27thinking | -2.10 |
| Benchmark | GPT-5 | Claude Opus 4 | Diff |
|---|---|---|---|
| ARC-AGI | 661 / 65 | 35.7048 / 65 | -29.70 |
| ARC-AGI-2 | 056 / 58 | 8.6038 / 58 | -8.60 |
| HLE | 6.30138 / 149 | 10.70121 / 149 | -4.40 |
| GPQA Diamond | 77.8081 / 175 | 79.6076 / 175 | -1.80 |
| Benchmark | GPT-5 | Claude Opus 4 | Diff |
|---|---|---|---|
| τ²-Bench | 8015 / 40thinking + 使用工具 | 72.5022 / 40thinking + 使用工具 | +7.50 |
| Benchmark | GPT-5 | Claude Opus 4 | Diff |
|---|---|---|---|
| SWE-bench Verified | 72.8041 / 103high | 72.5043 / 103 | +0.30 |
| Field | GPT-5 | Claude Opus 4 |
|---|---|---|
| Publisher | OpenAI | Anthropic |
| Release date | 2025-08-07 | 2025-05-23 |
| Model type | Foundation model | Reasoning model |
| Architecture | Dense | Dense |
| Parameters | 0.0 | Not available |
| Context length | 400K | 200K |
| Max output | 131072 | 32000 |
Prices use DataLearner records when available; missing fields are not inferred.
| Item | GPT-5 | Claude Opus 4 |
|---|---|---|
| Text input | 1.25 美元/100 万tokens | 15 美元/ 100万tokens |
| Text output | 10 美元/100 万tokens | 75 美元/100万tokens |
On average across the 11 shared benchmarks, GPT-5 scores 3.45 higher.
Largest single-benchmark gap: IMO-ProofBench — GPT-5 59 vs Claude Opus 4 2.90 (+56.10).
Page generated from structured model, pricing and benchmark records. No real-time LLM is used to write the prose.
