Anthropic launches Opus 4.7 and AI design tool, keeps Mythos private

Anthropic plans to release Claude Opus 4.7 and a generative design tool that builds websites, landing pages and slide decks from plain-English prompts. A person familiar with the plans confirmed the products could arrive as soon as this week. The company has not announced a release date publicly and did not respond to requests for comment.

The design tool targets both developers and nontechnical users and is intended to let customers prototype and deploy web pages and presentations without manual coding. Shares of Adobe, Wix and Figma fell on Monday after the plans became public.

Opus 4.7 is an update to Anthropic’s Opus line but is not the company’s most capable system. Anthropic is restricting access to Mythos, a version tuned for cybersecurity tasks, and is sharing it only with a small number of security firms.

A UK security research group evaluated a Claude Mythos preview and reported the model can plan and execute complex cyberattack simulations at higher rates than other public models. In a 32-step simulated corporate network assault called ‘The Last Ones,’ the group recorded Mythos completing the full exercise in three of ten attempts and averaging 22 successful steps out of 32. An earlier Opus 4.6 model averaged 16 successful steps. Human red teams typically require about 20 hours to complete the same simulation.

Anthropic develops its frontier systems by fine-tuning the Opus line, so improvements in Opus can later be hardened into specialized versions such as Mythos. Without a detailed model card or independent tests for Opus 4.7, comparisons with Mythos and other systems will rely on limited third-party evaluations and selective disclosures.

Industry evaluations produce varying results. Some recent tests showed low scores for top models on high-level reasoning tasks while human performance on those tests reached 100 percent. Companies and researchers continue to debate the reliability of benchmark results.

Recent company developments include leaked Claude code, the rollout of a skills system and an MCP protocol, and a greater focus on agentic AI and coding performance. The planned Opus 4.7 release and the AI design tool follow those product and enterprise developments.