Anthropic acquires Vercept to strengthen Claude computer-use capabilities
Original: Anthropic has acquired @Vercept_ai to advance Claude’s computer use capabilities. Read more: https://www.anthropic.com/news/acquires-vercept View original →
What Anthropic announced
Anthropic announced on February 25, 2026 that it acquired Vercept to push Claude's computer-use capabilities further. The company describes computer use as a way for Claude to operate inside live applications, enabling multi-step task execution across software environments that are difficult to solve through code-only interaction.
According to the announcement, Vercept was built around perception-and-interaction challenges for practical AI task completion. Anthropic named Vercept co-founders Kiana Ehsani, Luca Weihs, and Ross Girshick, and said the team will join Anthropic while winding down Vercept's external product in the coming weeks.
Performance context Anthropic provided
Anthropic linked the acquisition to its recent Claude Sonnet 4.6 release and highlighted progress on OSWorld, an AI computer-use evaluation. In the same post, Anthropic said its Sonnet models moved from under 15% in late 2024 to 72.5% at the time of announcement, and described current performance as approaching human-level on tasks like spreadsheet navigation and multi-tab form completion.
- Announced acquisition date: February 25, 2026.
- Stated strategic focus: computer-use reliability and capability depth.
- Reported benchmark movement: under 15% to 72.5% on OSWorld (company statement).
Why this deal is notable
The acquisition reflects a broader shift from pure chatbot UX toward agentic execution in real software environments. For enterprise buyers, the practical value is not only model quality in prompt-response tasks, but also stable operation across GUI workflows, context switching, and tool orchestration.
At the ecosystem level, this move also suggests that model labs are willing to combine internal model scaling with targeted capability acquisitions to close productization gaps. The next key indicators will be whether these computer-use gains hold under production reliability constraints and governance requirements across regulated workflows.
Related Articles
HNでの関心は、モデル名そのものより、同価格での改善、安くなったfast mode、Claude Codeのdynamic workflowsが実作業で効くかに集まった。
HNで注目されたのは「Claudeがバグを見つける」話だけでなく、各チームが自分の対象に合わせて作り替えるharnessの形だった。
Anthropicが出したのは単なる高性能モデルではなく、同じ基盤モデルを一般向けFableと限定向けMythosに分ける配布設計だ。価格は入力$10/出力$50、危険領域ではOpus 4.8への切り替えと30日保持も組み込まれる。