Anthropic Launches Claude Sonnet 5 and a Research Workbench, Putting Agent Power Into a Mid-Tier Model

On June 30, 2026, Anthropic released Claude Sonnet 5, its new mid-tier model, and set it as the default across Claude.ai, Claude Code, and the Claude Platform. The company described Sonnet 5 as the most agentic Sonnet it has ever shipped: the model can plan a task, drive a browser, run terminal commands, edit files, and recover from failed steps without a human stepping in. It is available to free, Pro, Max, Team, and Enterprise users, and is also being rolled out on Amazon Bedrock and Google Vertex AI.

The pricing tells the strategic story. Through August 31, the API costs two dollars per million input tokens and ten dollars per million output tokens, then rises to three and fifteen. Compared with Opus 4.8 at five and twenty-five, Sonnet 5 is roughly 60 percent cheaper during the launch window and about 40 percent cheaper afterward. Anthropic also updated the tokenizer, which means the same prompt now maps to between one and 1.35 times as many tokens, so heavy users should benchmark before assuming a flat bill.

The same week, Anthropic launched Claude Science, a research workbench that bundles more than sixty scientific databases and pre-built toolkits for genomics, protein structure, and chemistry. A main assistant breaks a research goal into steps, delegates them to sub-agents, and runs a fact-checker over the result. Every output carries a full provenance trail, so scientists can trace any claim back to the underlying data, and the environment runs on local macOS or Linux machines as well as over SSH to remote clusters.

The launches landed during a regulatory thaw. On July 1, the U.S. Commerce Department formally lifted export controls on Claude Fable 5 and Mythos 5, eighteen days after an order had forced Anthropic to take both models offline for any user, anywhere. The reversal followed Anthropic's addition of new safety classifiers that now block more than 99 percent of jailbreak attempts, and a written pledge to coordinate with the government on future model releases. Mythos 5 still opens only to a pre-approved list of U.S. organizations, while Fable 5 returns to global users in a phased rollout.