... there's more to Sonnet 4.5

Claude Sonnet 4.5 is being positioned as more than a faster, better coding model: it’s a stepping stone toward Anthropic’s “virtual collaborator,” an agent that can work on a user’s computer, execute tasks, and coordinate with other tools like Slack or Google Docs. The release matters because it targets the practical bottlenecks that keep coding assistants from becoming reliable long-running workers—agentic behavior, sustained focus, and tight integration with the user’s environment.

Early in the year, Anthropic’s CEO Dario Amodei previewed a 2025 plan for an agent that can take instructions, write and compile code, check its own work, and communicate progress back to the user while also interacting with co-workers’ systems. By the end of the third quarter, Sonnet 4.5 is framed as one component in that larger buildout. The model’s headline improvements include speed—early access users such as Devon reportedly saw it running about twice as fast as the prior model—while the bigger story shows up in benchmarks tied to agentic workflows rather than just raw coding quality.

On coding and agent-like tasks, Sonnet 4.5 is reported to outperform prior Claude versions and also to beat other major coding models on tests such as Sweetbench Verified, run across the full 500 EV vals rather than a reduced subset. A key detail is that the evaluation resembles “parallel test time compute,” a technique similar to what Gemini DeepThink uses; even without that, Sonnet 4.5 still shows a clear lead. For the virtual collaborator, coding is only the entry point. The model also needs to stay reliable over long stretches: Anthropic claims it can run up to 30 hours on complex tasks while maintaining focus, using an agent scaffold with multiple model calls.

Another requirement is direct interaction with the user’s tools. The computer-use benchmark is highlighted as a major jump over earlier Sonnet and Opus 4.1 versions, reflecting the model’s improved ability to operate in browser and desktop-like environments. That capability is paired with a major software layer: the Claude agent SDK. Renamed from the Claude Code SDK, it’s designed as a general agent harness rather than something locked to coding. The SDK emphasizes a loop that repeatedly gathers context, manipulates it (including via file-based memory), takes actions through built-in tools, custom tools, and MCPs, and then verifies outputs before iterating again.

Verification is treated as a differentiator, with examples including MCP-based tools like Playwright for screen-aware checks, plus LLM-as-judge methods to score and critique generated results. Beyond the SDK, Anthropic’s Claude Developer Platform adds backend context management—especially “context editing”—so long-running agents can summarize and compress older context while still preserving references through files. The practical impact is reinforced by Cognition’s Devon team, which reports faster and more reliable multi-hour sessions and notes that the model’s context-window behavior forced changes to agent architecture and prompting.

Finally, the virtual collaborator’s “eyes” are addressed through a Chrome extension for Mac plan users, with expectations it will expand to other paid tiers. With Sonnet 4.5, Anthropic appears to be assembling the model, the agent framework, and the context/control infrastructure needed to move from helpful assistant to operational collaborator—an approach aimed squarely at enterprise productivity rather than just developer demos.