Sonnet 5 Becomes Anthropic’s Flagship Model for Autonomous AI Tasks
Anthropic has released Claude Sonnet 5, calling it the most capable agent-focused Sonnet model the company has produced so far. Engineers designed it to draft plans, operate tools including web browsers and command-line terminals, and carry out extended autonomous work that once demanded larger, pricier systems. Compared with Sonnet 4.6, the new release delivers meaningful gains across reasoning, tool operation, software development, and general knowledge tasks, and it has already rolled out throughout Anthropic’s product line.
This launch carries weight because earlier Sonnet-tier models shaped Anthropic’s initial push into agentic AI. Versions 3.5, 3.6, and 3.7 of Claude Sonnet were among the first to demonstrate serious coding and tool-handling skill, though subsequent improvements in agentic ability had largely landed in the pricier Opus lineup instead. With Sonnet 5, Anthropic set out to close that divide, pushing capability near Opus 4.8 levels while holding the cost down.
The Model’s Strength Lies in Finishing Tasks, Not Just Conversing
According to Anthropic, the model targets agentic workloads specifically, work built around planning, reasoning across multiple steps, calling outside tools, and taking action over extended periods rather than answering a single prompt. The company frames Sonnet 5 as suited to jobs demanding continuous follow-through, professional-grade work included, at scale.
Across a set of effort levels, Sonnet 5 outperforms Sonnet 4.6 on every measured benchmark for agentic search and computer-use tasks, among them BrowseComp and OSWorld-Verified, without exception. Running at medium effort, it delivers a better balance of cost versus results, and pushed to higher effort, it can reach parity with Opus 4.8 on select tasks. Anthropic further states that the new model spans a broader spread of cost-to-performance tradeoffs than Sonnet 4.6 did.
This puts the model in a middle ground between raw capability and price. Opus 4.8 still stands as the broadly stronger model for comparison, yet Sonnet 5 running at its highest effort setting can equal Opus 4.8’s results on particular tasks.
Rollout Spans Every Tier, With Pricing Structured for Scale
As of today, Claude Sonnet 5 is accessible on every subscription tier. Free and Pro subscribers now get it as their default model, while Max, Team, and Enterprise customers can access it as well. Anthropic also made the model available within Claude Code and through the Claude Platform.
Launch pricing on the Claude Platform runs at $2 per million input tokens and $10 per million output tokens, a rate that holds through August 31, 2026. Once that window closes, the rate rises to $3 per million input tokens and $15 per million output tokens. By comparison, Opus 4.8 carries a price of $5 per million input tokens and $25 per million output tokens, meaning Sonnet 5 remains the cheaper option of the two.
Developers can reach the model through the Claude API by calling it claude-sonnet-5.
Internal Testing Points to a Safer, More Restrained Model
Anthropic reports that its internal safety testing turned up fewer instances of undesirable behavior in Sonnet 5 compared with Sonnet 4.6, making it, in the company’s view, a safer choice for agentic use cases overall. Separate evaluations, the company adds, found the model to be substantially weaker at cybersecurity-related tasks than its current Opus models.
These two findings together carry particular weight in agentic settings, where the system operates with greater independence, handling tools and executing actions without constant oversight. Anthropic points to this safety record as a reason to deploy Sonnet 5 in workflows that hand the model more operational responsibility.
Beta Users Report a Model That Pushes Through to Completion
Partners given early access reportedly delivered a unified message: Sonnet 5 behaves far more agentically than the Sonnet models before it. These testers noted that it carries complicated tasks through to completion in situations where prior Sonnet versions would have stalled, reviews its own work without prompting, and keeps progressing through jobs that used to require more hands-on intervention.
Where things stand now is simple to summarize: the model is live, its introductory pricing lasts until August 31, 2026, and it reaches consumers, teams, enterprises, and developers across all of Claude’s platforms. Anthropic characterizes this as putting agentic-level capability within reach at Sonnet’s price point, shrinking the gap between the mid-range Sonnet models and the costlier Opus tier.
For editorial consideration and industry coverage inquiries, contact [email protected]
