Anthropic launches Claude Sonnet 5 with 1M-token context, adaptive thinking on by default, and agentic-first design

Anthropic released Claude Sonnet 5 on June 30, 2026, as the next generation of its mid-tier Sonnet model family. Available immediately across all plans and via the Claude API under the model ID claude-sonnet-5, the release ships with introductory pricing of $2 per million input tokens and $10 per million output tokens through August 31, 2026, stepping up to $3 and $15 thereafter.

What's new

Sonnet 5 carries the same 1 million-token context window as its predecessor but raises the max output ceiling to 128,000 tokens. The headline capability change is that adaptive thinking is now on by default — the model autonomously adjusts its internal reasoning effort rather than requiring developers to configure it. Three API-breaking changes accompany the release:

• Manual extended thinking removed. The thinking: {type: "enabled", budget_tokens: N} parameter, deprecated on Sonnet 4.6, is gone entirely. Sending it now returns a 400 error.
• Sampling parameters locked. Setting temperature, top_p, or top_k to non-default values also returns a 400 error. Sonnet 5 manages its own sampling strategy.
• New tokenizer. The updated tokenizer produces approximately 30% more tokens for the same text, which means existing cost estimates and context budgets need recalculation.

Priority Tier inference is not available on Sonnet 5 — the model is excluded from that service tier.

On benchmarks, Anthropic reports Sonnet 5 as a strict improvement over Sonnet 4.6 on agentic evaluations including BrowseComp (long-horizon search) and OSWorld-Verified (computer use). Performance reportedly approaches Opus 4.8 at a fraction of the cost on many tasks.

The safety profile also improves: lower misaligned behavior rates, reduced hallucination and sycophancy, better refusal of malicious requests, and stronger resistance to prompt injection. The model launches with Anthropic's Cyber Verification Program safeguards enabled by default, and Anthropic notes it was intentionally undertrained on cybersecurity offensive tasks, resulting in substantially lower exploit development capability than Opus models.

Early enterprise partners cited the model's ability to finish complex tasks where previous Sonnet models would stop short, with particular strength in sustained coding and debugging workflows.

Context

Sonnet 5 follows Claude Sonnet 4.6, which had been Anthropic's workhorse API model since early 2026. The previous generation supported manual extended thinking and free sampling parameter configuration. The shift to locked sampling and mandatory adaptive thinking marks a deliberate architectural pivot: Anthropic is trading developer control for more consistent agentic performance.

The $2/$10 introductory pricing positions Sonnet 5 aggressively in the mid-market API tier before the August 31 step-up to standard rates.

Why it matters

The Sonnet tier is Anthropic's highest-volume API slot — the model most production deployments run on. Shipping the new Sonnet as an agentic-first model with adaptive thinking on by default signals that Anthropic now expects most Sonnet usage to be multi-step, tool-calling work rather than single-turn completions.

The removal of temperature and top_p control is the most immediately disruptive change for existing API users. Applications that tuned those parameters for creative or stochastic tasks will need to be rewritten — Sonnet 5 does not expose those knobs. Anthropic's bet is that adaptive thinking produces better outcomes than manual tuning across the workloads that matter.

The 30% tokenizer inflation is a quiet but significant cost change. An application that used 100K tokens per session on Sonnet 4.6 will use roughly 130K on Sonnet 5, all else equal, before the rate-card increase kicks in on September 1.