Claude Opus 5
Claude Opus 5 comes close to Fable 5 at half the price ($5/$25) and sets state of the art on Frontier-Bench and GDPval-AA.
Better at verifying and iterating on its own work. More than doubles Opus 4.8 on Frontier-Bench v0.1; 3x the next-best score on ARC-AGI 3; within 0.5% of Fable 5's peak on CursorBench 3.2 at half the cost. Cyber classifiers intervene about 85% less than on Fable 5. Default model on Max, strongest on Pro.
- Date
- Friday, 24 July 2026
- Lab
- Anthropic
- Kind
- model
- Access
- closed API
- Price
- $5/$25 per M tokens in/out, same as Opus 4.8; fast mode about 2.5x speed at 2x base price (2026-07-24)
Figures
| Measure | Value | Measured by |
|---|---|---|
| Frontier-Bench v0.1 | more than 2x Opus 4.8 also state of the art over all other models; at lower cost per task | company |
| ARC-AGI 3 | about 3x next-best model novel-problem solving | company |
| Organic chemistry (spectroscopy structure inference) | +10.2 points vs Opus 4.8 Anthropic internal benchmark; protein-sequence-effect tasks +7.7 | company |
| Automated behavioral audit, overall misaligned behavior | 2.3 lowest of recent models, below Opus 4.8, Sonnet 5 and Fable 5 | company |
| Terminal-Bench 4.0 | 52.3% Opus 5 as later reported in the Opus 5.5 table | company |
Not trained on cyber tasks yet close to Mythos 5 at finding vulnerabilities, but substantially behind it at developing exploits (OSS-Fuzz). Biology requests blocked on Fable 5 now fall back to Opus 5 instead of Opus 4.8. Launched with mid-conversation tool changes and automatic API fallbacks (beta). Benchmarks other than the ones quoted are in images; all figures are Anthropic or partner runs.
Sources
This record was checked against its sources on 6 October 2026. How we check