Claude Sonnet 5 brings agentic AI to more users

Anthropic has launched Claude Sonnet 5, an AI model designed to complete multistep tasks with less supervision and at a lower price than its Opus models. As of June 30, 2026, it is available on all Claude plans, as well as Claude Code and the developer API.

The difference lies in how it works. Sonnet 5 can create a plan, use a browser or terminal, review its own results, and continue until it completes a task. In practice, it does not just answer "how do I fix an error": it can investigate the problem, write a test, apply the change, and verify that the solution works.

Closer to Opus, at a lower cost

Anthropic says Sonnet 5 approaches the performance of Claude Opus 4.8 in tool use, coding, reasoning, and knowledge work. It also improves on Claude Sonnet 4.6, especially on tasks that require maintaining a plan across multiple steps.

Opus 4.8 remains the more accurate option in evaluations such as BrowseComp, which measures the ability to research with a browser, and OSWorld-Verified, which tests whether the model can operate computer interfaces. But Sonnet 5 offers a cheaper alternative for many everyday tasks.

That lets you choose between quality and cost based on how much effort each task requires. A simple automation can use fewer resources, while a complex investigation can give the model more time to reason.

What it can do in practice

The early-access tests cited by Anthropic describe improvements in tasks such as:

• Solving and verifying complex changes in existing codebases.
• Investigating bugs, reproducing them, and applying lasting fixes.
• Updating data in Salesforce and sending customer communications as part of the same task.
• Querying live data and producing analysis.
• Completing insurance workflows in applications companies already use.
• Carrying out multistep legal research and analysis.

The important promise is not that it writes better isolated answers. It is that it reaches the finish line more often on tasks that previously stalled, such as debugging a program, navigating multiple applications, or combining information from different sources.

Pricing and availability

Sonnet 5 is the default model for the Free and Pro plans. It is also available to Max, Team, and Enterprise users, as well as through Claude Code and Claude Platform.

In the API, the introductory price through August 31, 2026, is:

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

After that, the price will rise to $3 per million input tokens and $15 per million output tokens. Tokens are the units the model uses to divide text for processing.

There is one caveat: Sonnet 5 uses an updated tokenizer, the system that converts text into tokens. As a result, the same content may generate approximately 1 to 1.35 times more tokens, depending on the type of text. Anthropic says the introductory price is intended to make migration from Sonnet 4.6 approximately cost-neutral.

More capability does not eliminate the risks

Anthropic's safety evaluations found fewer hallucinations, less of a tendency to agree uncritically, and greater resistance to attempts to manipulate its instructions. It also recorded fewer undesirable behaviors than Sonnet 4.6 in an automated audit.

Even so, it is not the safest model in every scenario. Anthropic says Claude Opus 4.8 and Claude Mythos Preview achieved better overall results in some alignment tests. Sonnet 5 also showed somewhat greater capability than its predecessor on certain cybersecurity tasks, although it never managed to develop a complete exploit for a Firefox browser vulnerability in one of the evaluations.

For that reason, it ships with cybersecurity protections enabled by default. These safeguards detect and block dangerous uses in real time, such as attempts to create software that exploits vulnerabilities.

For you, the change depends on how you use Claude. If you only want answers or drafts, you will mainly notice access to the new model. If you connect it to tools, code, data, or applications, Sonnet 5 points to something more relevant: agents that not only know what to do, but can move forward on their own and deliver a verified result. The question to watch is whether that autonomy remains reliable when tasks become lengthy, ambiguous, or expensive.