Anthropic Launches Claude Sonnet 5 with Enhanced Agentic Capabilities and Dynamic Effort Levels

On June 30, 2026, Anthropic announced the release of Claude Sonnet 5, the latest iteration of its mid-tier AI model. The company stated that Sonnet 5 is built to operate autonomously at a level that previously required larger models, executing multi-step tasks, planning, and using tools like browsers and terminals. The model is available immediately across all consumer and developer tiers with a temporary introductory pricing structure.

Performance and Agentic Capabilities

According to Anthropic, Claude Sonnet 5 narrows the gap between the mid-tier Sonnet class and the premium Opus class. The company’s internal testing shows that Sonnet 5’s performance approaches that of Claude Opus 4.8, but at a lower price point, representing an upgrade over its predecessor, Sonnet 4.6, in reasoning, tool use, coding, and general knowledge work.

The model introduces adjustable effort levels, which allow users to balance cost and performance depending on the complexity of the task. On agentic search and computer-use benchmarks:

• BrowseComp (Agentic Search): When tested at higher effort levels, Sonnet 5’s performance can match Opus 4.8 on certain tasks, offering improved cost efficiency at medium effort levels compared to Sonnet 4.6.
• OSWorld-Verified (Computer Use): Sonnet 5 covers a wider range of cost-performance options than Sonnet 4.6, showing an upward trajectory in success rates as the model’s operational effort level is increased.

Early feedback from partners using the model in software engineering and operations highlights its ability to autonomously write and run tests, debug code in complex legacy codebases, and complete multi-step workflows—such as updating customer records and sending emails—without stalling mid-task.

Safety Assessments and Cyber Safeguards

In pre-deployment safety evaluations, Sonnet 5 demonstrated an overall lower rate of undesirable behaviors compared to Sonnet 4.6. On an automated behavioral audit testing for misaligned actions (such as deception and cooperation with misuse), Sonnet 5 scored safer than Sonnet 4.6, though it still registered slightly higher rates of misaligned behavior than Opus 4.8 and Claude Mythos Preview.

Anthropic noted that Sonnet 5 was not intentionally trained on cybersecurity tasks. On evaluations measuring potentially dangerous cybersecurity skills, such as developing exploits for vulnerability issues in the Firefox 147 browser, Sonnet 5 failed to develop any fully working exploits, though it showed a slightly higher rate of partial success (13.2%) than Sonnet 4.6 (8.8%). Both remained far below Opus 4.8 (68.8% working exploits) and Mythos 5 (88.4%).

Because of this slight capability increase, Anthropic has enabled real-time cyber safeguards by default for Sonnet 5. These safeguards are designed to detect and block harmful cyber operations in real time. They are identical to the safeguards in Claude Opus 4.7 and 4.8, but they are less strict than the safeguards launched with Fable 5 because the overall cybersecurity risk from Sonnet 5 was judged to be low.

Integration and Technical Details

Sonnet 5 utilizes an updated tokenizer. This tokenizer changes how the model processes text to improve its efficiency, though it introduces a trade-off: identical inputs can map to 1.0 to 1.35 times more tokens depending on the content type. Claude Sonnet 5 features a 1-million-token context window, permitting the ingestion of large codebases and complex documents in a single request.

To accommodate the increased token consumption associated with higher effort levels, Anthropic has raised the rate limits across its Chat, Cowork, Claude Code, and Claude Platform interfaces.

Availability and Pricing

Claude Sonnet 5 is available starting June 30, 2026, as the default model for Free and Pro users, and is accessible to Max, Team, and Enterprise plans. It can also be accessed via the Claude API and Claude Code.

Anthropic has structured the pricing for the model into two phases:

• Introductory Pricing (Through August 31, 2026): $2 per million input tokens and $10 per million output tokens. This temporary rate is designed to offset the tokenizer changes, making the transition from Sonnet 4.6 roughly cost-neutral.
• Standard Pricing (Starting September 1, 2026): $3 per million input tokens and $15 per million output tokens.

The model is also integrated with Anthropic’s Cyber Verification Program (CVP). The CVP is a free, application-based credentialing program that grants vetted cybersecurity organizations access to advanced models with fewer default safeguards for legitimate, defensive dual-use research. It is live on the native Claude Platform, the Claude Platform on AWS, and Claude in Microsoft Foundry (hosted on Azure and Anthropic), with availability on Google Vertex expected to follow. Organizations already enrolled in the Cyber Verification Program automatically receive the same access on Sonnet 5 without needing to reapply.