Claude Sonnet 5 brings Opus-like agent work to Free and Pro users

A cheaper model class just moved closer to the work that recently required Anthropic’s heavier Opus line. Claude Sonnet 5 is now the default model for Free and Pro users, and it is also available to Max, Team, Enterprise, Claude Code, and Claude Platform customers. The important shift is not only a model refresh. It is a wider rollout of browser, terminal, planning, coding, and long-running agent behavior into the tier many users actually touch every day.

Anthropic says Sonnet 5 improves over Sonnet 4.6 on reasoning, tool use, coding, and knowledge work. In the company’s agentic search and computer-use comparisons, the model opens a broader cost-performance range than the previous Sonnet generation and can approach Opus 4.8 on some higher-effort tasks. The commercial numbers matter: developers can call claude-sonnet-5 in the API at an introductory price of $2 per million input tokens and $10 per million output tokens through August 31, 2026. Standard pricing then moves to $3 input and $15 output per million tokens.

The launch also changes where agentic capability shows up operationally. Because Sonnet 5 is available in Claude Code and across paid team plans, it is positioned for software maintenance, test generation, pull request handling, data exploration, and multi-step business workflows rather than one-off benchmark demos. Anthropic’s examples from early users describe agents carrying messy coding tasks through reproduction, implementation, and verification. Those anecdotes are not independent evaluations, but they indicate the product target: more follow-through at a price that can support routine usage.

Safety is part of the release framing. Anthropic says pre-deployment tests found lower rates of undesirable behavior than Sonnet 4.6, better resistance to prompt injection, and cleaner refusals of malicious requests. The company also draws a boundary around cyber capability, saying Sonnet 5 is substantially weaker than Opus 4.8 and Mythos 5 on potentially dangerous cyber tasks. In a Firefox exploit evaluation, the Sonnet models did not produce a full working exploit, though Sonnet 5 showed a slightly higher partial-success rate than Sonnet 4.6. Cyber safeguards are enabled by default.

For builders, the practical question is whether Sonnet 5 makes agent workflows economical enough to run by default. The answer will depend on real workloads, tokenization changes, and how often higher-effort modes are needed. But the direction is clear: frontier-agent competition is moving from premium demos into default products and developer APIs.