Anthropic debuts Claude Sonnet 4.5 with Level 3 safety
Anthropic has introduced Claude Sonnet 4.5, a new version of its Claude family built to excel at coding, complex reasoning and hands-on computer use.
The model is aimed at building sophisticated agents and can operate directly within a browser—navigating sites, filling spreadsheets and completing multi-step tasks. Anthropic positions Sonnet 4.5 for use cases in finance, research and cybersecurity, where reliability and tool use matter.
Safety and release context
Anthropic released Sonnet 4.5 under its AI Safety Level 3 protections, which layer filters and classifiers to curb harmful inputs and outputs. The launch comes four months after Opus 4 and Sonnet 4, which debuted alongside the Level 3 framework, and lands amid rising enterprise interest in agentic AI. It also follows Microsoft’s decision to include Claude models, alongside OpenAI’s, in the 365 Copilot platform—raising Anthropic’s profile with business users.
Performance and benchmarks
Gartner analyst Arun Chandrasekaran said the release underscores Anthropic’s ambitions in coding performance as competition intensifies. To support that case, the company submitted Sonnet 4.5 to tests such as OSWorld (computer-use tasks), MMMLU for multilingual question answering, and MMMU for visual reasoning. Bradley Shimmin of Futurum Group added that Anthropic is targeting long‑standing issues like hallucinations and broader AI safety, areas he believes many market players still underweight.
New tools and early access features
Alongside the model, Anthropic rolled out the Claude Agent SDK to help developers build agentic applications. It also launched a research preview called Imagine with Claude, which the company says can generate software without requiring starter code. Imagine is available to subscribers on the Max tier for a limited five‑day window starting Monday.
Go‑to‑market questions
A key challenge, Chandrasekaran noted, is Anthropic’s route to market. The company has primarily reached enterprises through partners, and building a more direct motion could improve margins and monetization. Shimmin argued that, as a non‑hyperscaler, Anthropic may need to lean on partnerships to navigate a consolidating market—citing prior industry deals like Databricks’ 2023 acquisition of MosaicML. Anthropic’s approach resembles France-based Mistral AI, prioritizing a developer ecosystem for agentic tools and focusing revenue on inference rather than data capture.
Early integrations
Some vendors are already adding support. Glean said it will integrate Sonnet 4.5 into its no‑code agent builder.
What to watch
With enterprise demand for agentic capabilities growing, Sonnet 4.5’s real‑world performance, safety track record and the effectiveness of Anthropic’s partner strategy will determine how widely the model is adopted in the months ahead.
