Claude 4

Anthropic released Claude Opus 4 and Claude Sonnet 4 on May 22, 2025. The framing of the launch was a departure from the benchmark-first pattern of previous generations. Rather than leading with knowledge tests, Anthropic positioned the models around sustained performance on long-running tasks and agent workflows, the kind of work that unfolds over hours and thousands of steps rather than inside a single response.

The capability that made this possible was the ability to use tools during extended thinking. Earlier reasoning models produced an internal chain of thought and then acted; Claude 4 could alternate between reasoning and tool use within the same process, searching the web or running code partway through a thought and folding the result back into its reasoning. This closed a structural gap in agentic systems, where a model that could not gather evidence mid-reasoning had to guess or stop.

On coding evaluations, Claude Opus 4 reached 72.5% on SWE-bench and 43.2% on Terminal-bench, then the strongest published results on both. Pricing held at the level of the previous Opus and Sonnet generations, at fifteen dollars per million input tokens and seventy-five per million output tokens for Opus 4, and the models shipped simultaneously through Anthropic's API, Amazon Bedrock, and Google Cloud's Vertex AI.

Claude 4 shifted the unit of evaluation. For most of the preceding decade the question had been whether a model could produce a correct answer to a hard question. Claude 4 was built and marketed around a different question: how long can this system be given a goal, a set of tools, and no supervision before it drifts, loops, or quietly fails? Coding was the first domain where that question could be measured honestly, because a task either compiles and passes tests or it does not.

This reframing reorganized the engineering around models as much as the models themselves. Once a system could work unattended for an extended period, the surrounding infrastructure had to change to match: permissions governing what it could touch, sandboxes for code it wrote, audit trails of the actions it took, and evaluation suites that measured completed work rather than single responses. The agentic tooling that became standard across the industry over the following year was largely a response to capability arriving at this length.