Claude Sonnet 5: Anthropic's new model for agents and Claude Code

Anthropic announced Claude Sonnet 5 on June 30, 2026 in its official launch post and in the Claude Platform release notes. This is an important launch because Sonnet is the workhorse tier of Claude: cheaper and lighter than Opus, but powerful enough for coding, tool use and real product integrations.

The short version: Sonnet 5 brings much more agentic capability to the Sonnet family while keeping the price and availability profile closer to everyday production use. Anthropic says the model can plan, use tools such as browsers and terminals, and operate autonomously at a level that recently required larger and more expensive models.

This is not just a nicer chatbot release. Claude Sonnet 5 is primarily about work: code, tools, agents, long context and products built on the Claude Platform.

Claude Sonnet 5 at a glance

The main idea is simple: Claude Sonnet 5 aims to deliver much stronger agentic execution than Sonnet 4.6 without the full cost of Opus 4.8.

Why Sonnet 5 matters

Sonnet has never been merely Anthropic's middle model. For many developers, Claude's reputation in coding was built by Sonnet-class models. Claude Sonnet 3.5, 3.6 and 3.7 were cheap enough to use every day and strong enough to help with real code, refactors and repository analysis.

More recently, though, the largest agentic improvements have moved toward Opus and Fable-class models. Anthropic itself says that the clearest gains in planning, tool use and long-horizon work have been visible in Opus-class systems.

Claude Sonnet 5 is designed to rebalance that. It does not replace Claude Opus 4.8 as Anthropic's most capable general model, but it narrows the gap on important work while staying much cheaper. That matters for companies and builders who cannot afford to send every request to the most expensive model.

Claude Code and agentic coding

The most interesting context is Claude Code. Anthropic is not only describing a model that can write decent functions. It is describing a model that can maintain a plan, check its own work, use a terminal, debug and complete multi-step tasks without constant hand-holding.

That is exactly where developer tools are moving. Developers increasingly expect AI to do more than produce a snippet. The desired workflow is closer to: find the bug, write a reproducing test, implement the fix, check the regression and show the diff. If Sonnet 5 can do that at a lower cost than Opus, it could become the default model for everyday coding agents.

It is still worth being realistic. The hardest migrations and highest-stakes engineering work may still call for Opus or Fable. Sonnet 5 looks more like the model that should absorb the large middle layer of daily work: pull requests, bug fixes, smaller refactors, log analysis, tests, documentation and business automation.

Availability: where Claude Sonnet 5 can be used

Anthropic says Claude Sonnet 5 is available across Claude plans. It is the default model for Free and Pro users and is available to Max, Team and Enterprise users. It is also available in Claude Code and on the Claude Platform.

For developers, the key API identifier is:

claude-sonnet-5

That makes it straightforward to target the model directly. But teams migrating from Sonnet 4.6 should not simply change the model string and ship. The release notes include several changes that can affect existing integrations.

What developers need to know

Anthropic lists three migration changes that can break an integration if ignored.

First, adaptive thinking is on by default. The model decides how much reasoning effort to use based on the task. This replaces the older pattern where applications manually enabled extended thinking with a fixed token budget.

Second, manual extended thinking in this format:

{
  "thinking": {
    "type": "enabled",
    "budget_tokens": 10000
  }
}

does not work with Sonnet 5 and returns a 400 error. Anthropic notes that this path was already deprecated on Sonnet 4.6.

Third, setting non-default sampling parameters (temperature, top_p, top_k) also returns a 400 error. That matters for applications that previously tried to force more creative or more deterministic behavior through low-level model parameters.

There is one more practical detail: Claude Sonnet 5 uses a new tokenizer. Anthropic says the same text can produce roughly 30% more tokens. The introductory pricing is designed to make the transition roughly cost-neutral, but high-volume systems should still test real bills on real data.

Pricing: why the launch discount matters

Through August 31, 2026, Claude Sonnet 5 costs:

• $2 per million input tokens,
• $10 per million output tokens.

After that, the standard price becomes:

• $3 per million input tokens,
• $15 per million output tokens.

Compared with Opus 4.8, Sonnet 5 remains significantly cheaper. Anthropic gives Opus 4.8 as a higher-capability reference point priced at $5 per million input tokens and $25 per million output tokens.

In practice, Sonnet 5 may be most useful for teams building model routing. Cheaper models can handle simple tasks, Sonnet 5 can handle most difficult operations, and Opus or Fable can be reserved for the highest-stakes work.

Safety and cyber safeguards

Anthropic says Sonnet 5 has a lower overall rate of undesirable behaviors than Sonnet 4.6. The company also points to better refusal behavior for malicious requests and stronger resistance to prompt-injection hijacking attempts.

At the same time, Sonnet 5 is not being positioned as a top cyber-capability model. Anthropic says its performance on potentially dangerous cyber evaluations is substantially below Opus and Mythos. Even so, Sonnet 5 launches with cyber safeguards enabled by default. These safeguards are similar to those used in Opus 4.7 and 4.8, but less restrictive than the safeguards applied to Fable 5.

That distinction matters. Anthropic wants to offer a more capable Sonnet without carrying the full risk profile of its strongest models. After the recent events around Claude Fable 5, it is clear that the company will keep drawing sharper lines between everyday agentic models and systems with higher-risk capabilities.

Who should care about Sonnet 5

Claude Sonnet 5 looks most relevant for:

• developers using Claude Code;
• teams building AI agents;
• companies automating operational workflows;
• SaaS products that need a strong but not top-priced model;
• Claude users who want better performance than Sonnet 4.6 without moving everything to Opus.

It is not automatically the right model for every job. If the task is short and simple, a cheaper model may be enough. If the task is extremely difficult, long-running or high stakes, Opus 4.8 or Fable 5 may still be better.

That is why Sonnet 5 matters. The biggest shift in AI does not always come from the absolute strongest model. Sometimes it comes from the model that is good enough to use everywhere.

Verdict

Claude Sonnet 5 is one of Anthropic's most important 2026 launches because it targets the center of the market: daily coding, agents, automation and API usage. It is not trying to be the flashiest model in the portfolio. It is trying to be the model people actually use a lot.

If Anthropic's claims hold up in real workflows, Sonnet 5 could become the default choice for many Claude-based products: cheaper than Opus, more agentic than Sonnet 4.6 and strong enough to work seriously with tools.