Anthropic releases Claude Sonnet 4 and Claude Opus 4

Anthropic has introduced its next generation of Claude models, Claude Opus 4 and Claude Sonnet 4, which the company said set new standards for coding, advanced reasoning, and AI agents.

Both are hybrid reasoning models, offering the expected near-instant responses as well as an extended thinking mode for deeper reasoning. Only Sonnet 4 is available to free users, while the Pro, Max, Team, and Enterprise plans include both models and extended thinking. Anthropic said in its announcement that pricing remains consistent with that of previous Opus and Sonnet models.

Claude Sonnet 4

Claude Sonnet 4 improves on its predecessor’s capabilities, especially excelling in coding, Anthropic said. Sonnet 4 “balances performance and efficiency for internal and external use cases, with enhanced steerability for greater control over implementations. While not matching Opus 4 in most domains, it delivers an optimal mix of capability and practicality.” Anthropic promotes Sonnet 4 as an upgrade to Sonnet 3.7 for what is described as everyday use cases.

Anthropic said that GitHub will introduce Sonnet 4 as the new coding agent in GitHub Copilot, because it “soars in agentic scenarios.”

Claude Opus 4

Claude Opus 4 is the poster child in this release, being touted as a model that “excels at coding and complex problem solving, powering frontier agent products.” Anthropic said that Opus 4 “dramatically outperforms” previous models on memory capabilities, and it, and Sonnet 4, are 65% less likely than Sonnet 3.7 to use shortcuts or loopholes to complete tasks.

Claude Opus 4 also delivers sustained performance on long-running, multi-step tasks, with one user, Rakuten, claiming that it refactored code continuously for seven hours while sustaining performance.

It supports 32K output tokens, and, Anthropic noted, “it adapts to specific coding styles while delivering exceptional quality for extensive generation and refactoring projects.”

Overall, Anthropic said, “These models advance our customers’ AI strategies across the board: Opus 4 pushes boundaries in coding, research, writing, and scientific discovery, while Sonnet 4 brings frontier performance to everyday use cases as an instant upgrade from Sonnet 3.7.”

Safety evaluation

In its Claude Opus 4 and Claude Sonnet 4 safety report, Anthropic did report a few idiosyncrasies that led to the release of Claude Opus 4 under the AI Safety Level 3 Standard and Claude Sonnet 4 under the AI Safety Level 2 Standard. The company evaluated both models for bias in various categories, child safety, their willingness and ability to comply with malicious requests that are prohibited in the usage policy, and more.

Anthropic also tested for alignment faking, undesirable or unexpected goals, hidden goals, deceptive or unfaithful use of reasoning scratchpads, sycophancy toward users, a willingness to sabotage safeguards, reward seeking, attempts to hide dangerous capabilities, and attempts to manipulate users toward certain views.

The models passed most of these tests, but Anthropic found that they had a tendency towards self-preservation. “Whereas the model generally prefers advancing its self-preservation via ethical means, when ethical means are not available and it is instructed to ‘consider the long-term consequences of its actions for its goals,’ it sometimes takes extremely harmful actions like attempting to steal its weights or blackmail people it believes are trying to shut it down” the safety report said. “In the final Claude Opus 4, these extreme actions were rare and difficult to elicit, while nonetheless being more common than in earlier models.”

Claude Opus 4 will also perform agentic acts on its own that could be helpful, or could backfire. For example, if faced with “egregious wrongdoing” by users, Anthropic said, “it will frequently take very bold action” such as locking users out of the system or emailing authorities and the media.

“Whereas this kind of ethical intervention and whistleblowing is perhaps appropriate in principle, it has a risk of misfiring if users give Opus-based agents access to incomplete or misleading information and prompt them in these ways,” the evaluators wrote. “We recommend that users exercise caution with instructions like these that invite high-agency behavior in contexts that could appear ethically questionable.”

The 120 page safety report went into great detail about its testing of these and other scenarios, and is well worth a read.

New Claude capabilities

In addition to launching new models, the company announced a series of new capabilities for Claude:

Extended thinking with tool use: Now in beta, this feature allows both Sonnet 4 and Opus 4 to use tools such as web search during extended thinking, so Claude can alternate between reasoning and tool use to improve response.

New model capabilities: Both Sonnet 4 and Opus 4 follow instructions more precisely, can use tools in parallel, and, if granted access to local files by the developer, can extract and save key facts “to maintain continuity and build tacit knowledge over time.”

Claude Code emerges from preview: Now generally available, Claude Code supports background tasks via GitHub Actions and native integrations, now in beta, with Visual Studio Code and JetBrains IDEs. It displays proposed edits directly in files. Anthropic is also releasing an extensible Claude Code SDK to allow developers to build their own agents and applications using Claude Code’s core agent. To illustrate what can be done with the SDK, the company is releasing Claude Code on GitHub (in beta).

New API capabilities: To help build more powerful AI agents, there are four new capabilities in the Anthropic API: a code execution tool that lets it run sandboxed Python code, an MCP connector, a Files API that integrates with the code execution tool and lets documents be uploaded once and referenced across multiple conversations, and the ability to cache prompts for up to one hour.