29 June – 5 July 2026

The document details technical specs and safety testing for the model, which the company identifies as its most agentic Sonnet release to date. The System Card arrives alongside the general availability of Sonnet 5 for all users via the Claude interface and API.

The report covers model behavior in agentic scenarios, safety evaluations, and resilience against misuse. At the Extra High setting, Sonnet 5 matches the performance of Opus 4.8 in standard mode across the OSWorld-Verified and BrowseComp benchmarks.

The full document is available on the Anthropic website.

The /rewind command now functions even after a /clear. If context was cleared by mistake, users can roll back to the point where the data remained intact.

/usage now breaks down tokens by category: code, history, and specific tools, replacing the previous single counter. The Claude agents panel displays all parallel background sessions filtered by the initial prompt URL, and workers now auto-recover after a daemon restart.

The fallbackModel parameter allows setting up to three backup models. If the primary is overloaded, Claude Code switches automatically and retries the request. The /code-review command has been consolidated from five separate passes into one, saving approximately 25% in tokens. Finally, /mcp now displays servers with zero connected tools, making it easier to identify connection failures.

According to Reuters, Alibaba is prohibiting employees from using Claude Code, citing potential backdoor risks. The formal trigger is code Anthropic added to Claude Code in March to identify competitors.

That code checked for base URL overrides and, when connected via proxy, compared the timezone and host against a list of Chinese AI labs and resellers. Anthropic states it removed this logic in the July 1 release.

The ban follows a broader conflict. Anthropic previously accused Alibaba of unauthorized access to Claude via thousands of fraudulent accounts linked to the Qwen lab. Both parties are now blocking each other's tools within their respective perimeters.

Anthropic has opened a Cyber Jailbreak Vulnerability Disclosure Program on the HackerOne platform. Security researchers are invited to identify and report vulnerabilities that allow users to bypass the safety mechanisms of its models.

The company uses HackerOne to locate and remediate critical security flaws before they can be exploited by malicious actors.

Anthropic plans to develop its own drugs rather than just providing AI tools for pharmaceutical companies, according to a report from STAT News.

This move follows the launch of Claude Science, a workspace built for life science researchers. Anthropic is now shifting from a tool provider to an active participant in drug development.

Specific program details and therapeutic focus areas have not yet been disclosed.