On June 30, 2026, Anthropic released Claude Sonnet 5, which it calls its most agentic Sonnet model yet: one that can plan, use tools such as browsers and terminals, and work autonomously at a level that previously required larger, more expensive models. Anthropic says Sonnet 5 performs close to Claude Opus 4.8 while costing $2 per million input tokens and $10 per million output tokens, less than half of Opus 4.8's price. For developers, the release changes the default cost-performance calculation for agents, and it brings a new tokenizer that affects how many tokens the same text consumes.

Key Facts#

  • Released: June 30, 2026, on the Claude API as claude-sonnet-5 and across Anthropic's plans and platforms.
  • Default model: Sonnet 5 became the default for Free and Pro plans and is available to Max, Team and Enterprise users.
  • Pricing: $2 per million input tokens and $10 per million output tokens. Anthropic launched this as introductory pricing and, on August 10, 2026, made it permanent, cancelling a planned move to $3 and $15 on September 1.
  • Positioning: Anthropic says Sonnet 5 is close to Opus 4.8, which costs $5 and $25 per million tokens, and substantially better than Sonnet 4.6 on reasoning, tool use, coding and knowledge work.
  • Tokenizer: a new tokenizer can map the same input to roughly 1.0 to 1.35 times as many tokens, depending on content type.
  • Safeguards: cyber safeguards are enabled by default, the same ones used for Claude Opus 4.7 and 4.8, and less strict than those on Claude Fable 5.
  • Rate limits: Anthropic raised limits across Chat, Cowork, Claude Code and the Claude Platform to accommodate heavier use of higher effort levels.

What Happened#

Anthropic framed Sonnet 5 as a return to form for its mid-tier model family. It noted that Sonnet models, from Claude Sonnet 3.5 through 3.7, were the first to show strong coding and tool-use skills, but that the clearest recent gains in agentic capability had come from the larger Opus models. Sonnet 5 narrows that gap. Early-access testers told Anthropic that it finishes complex tasks where earlier Sonnet models stopped short, and that it checks its own output without being asked.

The launch emphasized effort levels rather than a single benchmark score. Anthropic published cost-performance curves for the BrowseComp agentic search benchmark and the OSWorld-Verified computer use benchmark. Anthropic described Sonnet 5 as a strict improvement over Sonnet 4.6, offering a wider range of cost and quality trade-offs than Opus 4.8 and matching Opus 4.8 on some tasks at higher effort. Anthropic corrected its BrowseComp chart on launch day, saying the original used a simpler methodology that underestimated Sonnet 5. As always, these are vendor-reported results.

Safety testing found Sonnet 5 better than Sonnet 4.6 on refusing malicious requests, resisting hijacking through prompt injection, and avoiding hallucination and sycophancy. It showed somewhat more misaligned behavior than Opus 4.8 and Claude Mythos Preview on Anthropic's automated audit. Anthropic said it did not deliberately train Sonnet 5 on cybersecurity tasks. On an evaluation built with Mozilla to test exploit development for Firefox vulnerabilities, it never produced a full working exploit, although its rate of partial success rose slightly over Sonnet 4.6. Because the model is somewhat stronger here than its predecessor, Anthropic enabled real-time cyber safeguards by default and made Sonnet 5 part of its Cyber Verification Program.

Background#

Sonnet 5 replaced Claude Sonnet 4.6, released on February 17, 2026, which was also the default on Free and Pro plans and was priced from $3 and $15 per million tokens. Sonnet 4.6 had been a strong step itself: Anthropic said early users often preferred it even to Claude Opus 4.5, its top model from November 2025. The Sonnet line is typically where developers balance price and performance, as our coverage of Claude Sonnet 4.5 described.

The launch came three weeks after Anthropic released Claude Fable 5 and Mythos 5 at the top of its lineup, and after an April 26 change that raised Sonnet and Haiku rate limits at every usage tier and simplified the native Claude Platform to three tiers: Start, Build and Scale. With Sonnet 5, Anthropic filled the middle of its lineup with a model that makes agentic workflows cheaper.

Why It Matters for Developers#

For many teams, a Sonnet-class model is the one that runs at volume, so the details matter more than the headline.

  • Re-baseline costs after the tokenizer change. The same prompt can cost up to 35% more tokens than on Sonnet 4.6, which partly offsets the lower price and shrinks the effective context window. Re-measure token counts, cost per request and context budgets with real traffic instead of reusing old estimates. Our LLM observability and cost guide shows how to capture this with OpenTelemetry.
  • Route by task, not by habit. With near-Opus quality at a lower price, a sensible default is Sonnet 5 for most calls, escalating to an Opus model only for tasks your evaluations show need it. Coding against IChatClient from Microsoft.Extensions.AI keeps that routing in configuration, and our guide to AI agent architecture patterns covers router designs.
  • Tune effort per feature. Effort levels trade cost and latency for quality. Pick them per use case and validate each choice with your own test set, as described in our guide to evaluating AI applications.
  • Expect some security prompts to be blocked. Real-time cyber safeguards are on by default. Teams building legitimate security tooling should test their prompts, consider the Cyber Verification Program, or follow Anthropic's recommendation to use Opus 4.8 for work that needs reduced guardrails.

Better resistance to prompt injection is welcome, but it is not a security boundary. Keep least-privilege tools and output validation in place, as our guide to responsible AI and LLM security recommends.

What's Next#

Anthropic's lineup kept moving quickly. On July 24, it released Claude Opus 5. On August 10, it made Sonnet 5's introductory pricing permanent. On September 22, it released Claude Opus 5.5, whose announcement said Claude Sonnet 5.5 and Claude Haiku 5.5 would follow in the coming weeks. At that point, Anthropic had not said how Sonnet 5.5 will be priced or whether it keeps Sonnet 5's tokenizer.

Sources#