Claude Opus 4.8

Claude Opus 4.8 is the most capable model Anthropic currently ships. It is built for work where accuracy and depth matter more than speed or cost, and it is the model most teams reach for when a task involves long chains of reasoning, large codebases, or research questions that smaller models get almost right but not quite.

The headline change over the earlier Opus 4 is context. Opus 4.8 reads up to one million tokens in a single request, which is roughly a large repository or several books at once, so it can hold an entire project in view instead of working from fragments. Paired with extended thinking, where the model reasons through a problem before it answers, this makes it well suited to multi step planning, architectural decisions, and debugging that spans many files.

On public benchmarks Opus 4.8 leads the family. It scores in the mid ninety percent range on GPQA Diamond, a test of graduate level scientific reasoning, and posts the strongest SWE bench result of any Claude model so far, which measures how often a model resolves real software issues end to end. The numbers matter less than the pattern behind them. The model tends to fail softly, asking for clarification or flagging uncertainty rather than inventing an answer.

Opus 4.8 is priced as a frontier model, so it is not the right default for high volume classification or simple chat. For those jobs Claude Haiku 4.5 is far cheaper and fast enough. Many teams route the bulk of their traffic to a smaller model and escalate only the hardest requests to Opus 4.8, which keeps costs under control without giving up quality where it counts.

Where does it sit next to the newer Claude 5 generation? Claude Fable 5 is the creative and long form specialist and often reads better on open ended writing, while Opus 4.8 remains the reference point for rigorous reasoning and agentic coding. For security research and scientific discovery, Anthropic keeps the gated Claude Mythos model, which is not generally available.

If you are choosing a model today, start with the question you are trying to answer. Reasoning heavy work, large context and autonomous coding all point to Opus 4.8. You can compare the whole lineup on the Claude model family page, or read our coverage of the Opus benchmark results for the numbers in context.

Cost control on Opus 4.8 comes down to two habits. Prompt caching lets you reuse a large fixed context, such as a codebase or a policy document, without paying full input rates on every call, which matters when the context runs to hundreds of thousands of tokens. Extended thinking is billed at input rates, so setting a sensible thinking budget keeps spend predictable while still giving the model room to plan. Used together, they make the frontier model affordable for the requests that genuinely need it, and wasteful only if you point it at work a smaller model could handle.

Opus 4.8 is available through the Anthropic API, the Claude apps and the major cloud platforms. Extended thinking, prompt caching and vision are all supported, and the same model powers both the assistant and the developer tools, so behaviour stays consistent from prototype to production.