Claude Sonnet 5 - New Agentic AI Model

Claude Sonnet 5 is now available as the latest agentic model from Anthropic.

It runs autonomously using browsers and terminals.

Consequently, it performs tasks that previously required larger artificial intelligence models.

This release narrows the performance gap with the larger Opus 4.8 model while maintaining lower operational costs.

Specifically, it improves upon Sonnet 4.6 in reasoning, coding, and knowledge work.

In addition, the model uses an updated tokenizer that changes how it processes text to improve performance.

Capabilities of Claude Sonnet 5

Developers can access the model through the Claude API.

Meanwhile, the model has been integrated as the default option for Free and Pro plans.

It is also available to Max, Team, and Enterprise users.

Furthermore, it is available in Claude Code and on the Claude Platform.

Safety assessments show that Claude Sonnet 5 has a lower rate of undesirable behaviors compared to Sonnet 4.6.

However, its cybersecurity capabilities remain limited.

As a result, the company has enabled default cyber safeguards to detect and block dangerous usage in real time.

During evaluations on Firefox 147 vulnerabilities, the model was unable to develop a full working exploit.

Notably, it showed a slightly higher rate of partial success than Sonnet 4.6 due to general intelligence improvements.

The model is part of the Cyber Verification Program on AWS, Microsoft Foundry, and Google Vertex.

Organizations already enrolled in this program receive automatic access.

Early access partners stated that the model is more agentic than previous versions.

Specifically, testers noted that it finishes complex tasks.

Furthermore, it checks its own output without explicit instructions from the user.

The model launches with introductory pricing of $2 per million input tokens and $10 per million output tokens.

This promotional rate is active through August 31, 2026.

After this date, standard pricing will be $3 per million input tokens and $15 per million output tokens.

On the Humanity’s Last Exam benchmark, the previous model scored 34.6% without tools and 46.8% with tools.

Meanwhile, on the OSWorld-Verified evaluation, the previous model scored 78.5% after evaluation adjustments.

The new model demonstrates strict improvements across these benchmarks.

Rate limits have been increased across Chat, Cowork, Claude Code, and the Claude Platform.

Therefore, developers can select different effort levels to balance cost and accuracy.

Ultimately, Claude Sonnet 5 provides a lower-priced option of higher quality.

This allows users to adjust effort levels to find the right balance of cost and performance.