Claude Opus 4.8
Last updated: August 10, 2026
Specs
| Vendor | Anthropic |
|---|
| Released | May 28, 2026 |
|---|
| Context window | 1M tokens |
|---|
| Input | $5 / M tokens |
|---|
| Output | $25 / M tokens |
|---|
Benchmarks
| mmlu | 93 |
|---|
| gpqa | 67.5 |
|---|
| humanEval | 94.5 |
|---|
| arenaElo | 1435 |
|---|
| sweBench | 67 |
|---|
| sweBenchPro | 69.2 |
|---|
| liveCodeBench | 78 |
|---|
| aiderPolyglot | 82 |
|---|
| terminalBench | 50 |
|---|
| mmluPro | 87 |
|---|
| hle | 24 |
|---|
| aime | 90 |
|---|
| math500 | 97 |
|---|
| mmmu | 80 |
|---|
Strengths
- Leads agentic coding: 69.2% SWE-Bench Pro, ahead of Opus 4.7 (64.3%), GPT-5.5 and Gemini 3.1 Pro
- Strongest computer-use / browser-agent tested: 84% on Online-Mind2Web
- Longer autonomous runs and more honest self-reporting; 1M-token context with 128K max output
Weaknesses
- Frontier-tier list price, overkill for high-volume, latency-sensitive traffic
- Slower than Haiku/Sonnet on simple tasks unless run in fast mode
Best for
- Autonomous engineering agents
- Browser and computer-use automation
- Long-running research and document analysis
Compare with other AI models