Claude Sonnet 5 is Anthropic’s AI model for coding, agentic tasks, and professional work. It is designed to balance capability, speed, and cost. Anthropic released it on June 30, 2026, and it is available in Claude products and through the Claude API as claude-sonnet-5.
What is Claude Sonnet 5?
It is the fifth-generation Sonnet model and a direct upgrade from Claude Sonnet 4.6. It supports a one-million-token context window, up to 128,000 output tokens, text, images, PDFs, and tool use.
What makes it different?
Sonnet 5 focuses on coding, planning, and multi-step execution. It supports adaptive thinking, prompt caching, batch processing, the Files API, PDF support, vision, and server-side and client-side tools. Compatible deployments also support computer use and browser use.
How does it compare with Claude Opus 5?
Opus 5 is the higher-capability tier for the most demanding work. Sonnet 5 targets a more practical balance of speed, price, and intelligence. It is therefore better suited to daily coding, scalable agent workflows, and high-volume professional tasks, while Opus is intended for work that needs Anthropic’s stronger premium model.
How much does Claude Sonnet 5 cost?
The current standard price is $2 per million input tokens and $10 per million output tokens. Cache hits and refreshes cost $0.20 per million tokens, while cache writes have separate rates. Anthropic confirmed in August 2026 that the introductory $2/$10 pricing became permanent.
What are its limitations?
The model uses a new tokenizer that may produce roughly 30% more tokens for the same text than Sonnet 4.6. Lower per-token pricing therefore does not always translate into an equivalent reduction in total task cost.
Non-default sampling parameters such as temperature, top_p, and top_k are not accepted. The older manual extended-thinking configuration has also been replaced by adaptive thinking.
Real-time cybersecurity safeguards may refuse some high-risk requests. This improves safety but can restrict legitimate security work unless an organization qualifies for Anthropic’s Cyber Verification Program.
What did its safety evaluation show?
Anthropic reports that Sonnet 5 achieved a 0.0% full-working-exploit success rate in a Firefox vulnerability evaluation run without safeguards. This is a narrow cybersecurity safety result, not a general measure of coding quality.
In short
Claude Sonnet 5 is a practical model for coding, agents, and knowledge work. It combines long context and strong tool support at a lower price than the Opus tier, but migrations should account for tokenizer changes, API constraints, and cybersecurity safeguards.