Claude Sonnet 5: Anthropic's Most Agentic Model Becomes the New Standard

On June 30, 2026, Anthropic launched Claude Sonnet 5, available in the API under the identifier claude-sonnet-5. The company describes it as “the most agentic Sonnet yet” — a model that doesn’t merely answer questions but plans, executes, and self-evaluates multi-step tasks using tools such as web browsers and terminals. On select task benchmarks, Anthropic claims Sonnet 5 approaches the performance of Opus 4.8 while carrying a significantly lower price tag.

Claude Sonnet 5 Immediately Becomes the Default Model Across All Plans

Starting today, Sonnet 5 replaces its predecessor as the default model on Free and Pro plans. It is also available on Max, Team, Enterprise, and Claude Code subscriptions. On the same day, Claude Code v2.1.197 sets Sonnet 5 as the new default model in the development environment, giving development teams a 1 million token context window — equivalent to roughly 750,000 words of text or an entire large codebase — together with 128,000 tokens of maximum output per call.

This combination of a large context window and agentic capabilities is particularly relevant for teams working on long code-base analysis, complex documentation, or sustained autonomous tasks where context continuity is critical.

Introductory Pricing and a New Tokenizer

Anthropic is introducing an aggressive introductory price valid through August 31, 2026: $2 per million input tokens and $10 per million output tokens. That is 33% below the standard pricing that takes effect in September: $3 input / $15 output per MTok.

Developers migrating from Sonnet 4.6 need to account for one important technical change: Sonnet 5 uses a new tokenizer shared with Opus 4.7 and Fable 5. Depending on content type, the same text may tokenize into up to 35% more tokens than on Sonnet 4.6. For high-volume applications, this tokenization difference could partially offset savings from the lower intro price — testing with real traffic is recommended before making a definitive cost assessment.

What Changes in the API for Developers?

The move to Sonnet 5 brings three breaking changes relative to Sonnet 4.6 that can silently break existing integrations:

1. Adaptive thinking is enabled by default. Unlike Sonnet 4.6, where this option had to be explicitly activated, Sonnet 5 uses extended thinking automatically when it judges it useful for a task. Applications that assumed deterministic output without a thinking block will need to update their response parsing.

2. The manual budget_tokens parameter has been removed. The API in Sonnet 4.6 accepted budget_tokens to cap thinking explicitly. In Sonnet 5, sending this parameter returns an HTTP 400 error. Every API call that explicitly sends it must be updated before migration.

3. Non-standard sampling parameter values return error 400. Values of temperature, top_p, and top_k outside recommended ranges now immediately return HTTP 400. Experimental implementations that used extreme values of those parameters will need to be adjusted.

The recommended approach for teams: run the new model in a parallel staging environment, validate all API calls with status-code logging, then migrate production.

Agentic Capabilities and Performance

Sonnet 5’s core differentiator is the ability to plan and execute multi-step tasks. The model can devise a plan for a complex goal, use tools — browser, terminal, file system — to carry it out, then evaluate whether the goal has been achieved and correct course if needed. This makes it well-suited for automating development tasks, research, report generation, and workflows where previous models could not autonomously complete the work.

Anthropic reports substantial improvements over Sonnet 4.6 on reasoning, tool use, coding, and complex problem-solving benchmarks. On certain evaluations, performance is statistically indistinguishable from Opus 4.8 — which is especially significant given the price difference between the two models.

Safety Improvements and Cyber Profile

Sonnet 5 brings measurable improvements in model reliability and safety. It exhibits fewer hallucinations and less sycophancy than Sonnet 4.6, is more resistant to prompt injection attacks, and better refuses malicious requests. Cybersecurity guardrails are active by default, and Anthropic emphasizes that the model has substantially lower cyber-offensive capabilities than Opus models — a deliberate decision to differentiate the safety profiles of models at different capability levels and use cases.

Amazon Bedrock and Broader Ecosystem Availability

Claude Sonnet 5 launched simultaneously across multiple platforms. On Amazon Bedrock it is available under the identifier us.anthropic.claude-sonnet-5, and Anthropic positions it as “frontier-grade intelligence at Sonnet pricing” for users of AWS infrastructure. It is also available on the Claude Platform on AWS at identical introductory prices.

The simultaneous cross-platform rollout — Anthropic directly, Amazon Bedrock, and GitHub Copilot (where Sonnet 5 also reached general availability on the same day) — marks a coordinated launch strategy that delivers a consistent experience regardless of infrastructure platform. For teams considering migration: API compatibility testing, particularly around the three breaking changes described above, should be a day-one priority.