Claude Opus 4.8
Claude Opus 4.8 by Anthropic: hybrid reasoning model with adaptive thinking, effort control, four times fewer unremarked code flaws, and Mythos-level alignment scores.
Model Overview
Capabilities, design details, and architectural traits
Claude Opus 4.8 - Hybrid Reasoning Model Built for Sustained Agentic Work
Claude Opus 4.8 is Anthropic's most capable publicly available model prior to the Fable launch, officially described as a hybrid reasoning model built for serious coding and AI agents. Its defining design principle is adaptive thinking - the model automatically adjusts how much reasoning it applies based on task complexity, spending more computation on harder problems and responding faster to simpler ones. This is always on and user-controllable via effort settings (including an xhigh tier for maximum computation).
What Distinguishes Opus 4.8 From Its Predecessors
| Trait | What it means for Opus 4.8 specifically |
|---|---|
| Adaptive thinking, always on | Automatically scales reasoning depth per task; users can also set effort explicitly from low to xhigh |
| Four times fewer unremarked code flaws | Officially documented: Opus 4.8 is around four times less likely than Opus 4.7 to allow flaws in its own code to pass without comment |
| Alignment scores matching Mythos Preview | Evaluation-confirmed rates of deceptive behavior and cooperation with misuse are similar to Claude Mythos Preview, Anthropic's most aligned model at launch |
| Fast mode at 2.5x speed, 3x cheaper than prior models | Fast mode runs at 2.5 times the speed and costs three times less than fast mode did on previous Opus versions |
| 1M token context window | Supports a 1 million token context window for sustained, long-running sessions |
| Fallback target for Fable 5 safety classifiers | When Claude Fable 5 declines a flagged request, the API automatically reroutes to Opus 4.8 - a documented platform role unique to this model |
Honesty as a Documented Behavioral Shift
Anthropically's official announcement frames honesty as the most prominent improvement in Opus 4.8 over 4.7. The model is explicitly trained to flag uncertainties rather than assert unsupported progress - particularly relevant in agentic coding sessions where overconfident claims about task completion are a documented failure mode in prior models.
Benchmark Performance
Independent evaluations · Artificial Analysis
Accuracy & Capability Details
Independent evaluation data provided by Artificial Analysis. To view the latest benchmarks and full details, visit their official site.
Compare Models Side-by-Side
Evaluate specifications, pricing, and independent benchmark indices
| Model Details | AnthropicHost Claude Opus 4.8 | ||
|---|---|---|---|
| General Info | |||
| Provider | Anthropic | Anthropic | Anthropic |
| Release Date | May 28, 2026 | June 9, 2026 | April 16, 2026 |
| Knowledge Cutoff | — | — | — |
| Context & Limits | |||
| Context Window | 1M | 1M | 1M |
| Pricing (per 1M tokens) | |||
| Input Pricing | $5 Best Input Pricing | $10 | $5 Best Input Pricing |
| Output Pricing | $25 Best Output Pricing | $50 | $25 Best Output Pricing |
| Modalities | |||
| Inputs | textimagefile | textimagefile | textimagefile |
| Outputs | text | text | text |
| Benchmarks (0-100) | |||
| Intelligence Index | 55.7 | 59.9 Best Intelligence Index | 53.5 |
| Coding Index | 74.3 | 76.5 Best Coding Index | 73.6 |
| Agentic Index | 47.2 | 52.8 Best Agentic Index | 44.4 |
GPQA Benchmark
Graduate-level reasoning and expert Q&A evaluation.
Humanity's Last Exam
Extremely difficult logical reasoning and knowledge.
IFBench Instruction
Evaluation of strict instruction following.
Benchmark scores are independent evaluations sourced from Artificial Analysis. Intelligence, Coding, and Agentic indices reflect composite performance ratings (0–100).