Anthropic has launched a new version of chatbot Claude. In addition to improved performance, the model also proves to be effective at detecting vulnerabilities.
Anthropic has launched Claude Opus 4.6, a new version of its language model that scores better on software development, information retrieval, and complex reasoning tasks. The model now features a context window of 1 million tokens in beta. Claude Opus 4.6 also proves to be an excellent bug hunter.
Stronger performance in software and knowledge work
Claude Opus 4.6 offers improvements in planning, debugging, and coding within larger software projects. According to benchmarks shared by Anthropic, the model scores highest on autonomous coding and complex reasoning, and demonstrates strong knowledge of economics and law.
The new Claude model is also more effective at finding hard-to-locate information and remains accurate during lengthy tasks by better retaining contextual information. In tests with long texts, Opus 4.6 scored 76 percent on the MRCR v2 benchmark, where previous versions still achieved a clear fail.
Within the Claude Code development environment, users can now assemble agent teams that work on tasks in parallel. Through the API, Anthropic introduces context compaction, adaptive reasoning capability, and an output limit of 128,000 tokens per task.
Broader deployment for professional work
Claude Opus 4.6 gains broader applications for daily office work. Claude in Excel has been expanded with capabilities to process unstructured data and perform multi-step operations. Claude in PowerPoint, currently in testing phase, automatically reads formatting settings and assists in creating presentations based on data from Excel.
read also
Coding with a Single Prompt? ‘Vibe Coding’ is No Miracle Cure
500 vulnerabilities
Anthropic states that Claude can also serve software security. To demonstrate this, Anthropic unleashed Claude on open source software. Claude Opus 4.6 discovered no fewer than 500 previously unknown CVEs.