Claude Sonnet 5 is here: more affordable and more affordable

Anthropic is updating its mid-range AI model: Claude Sonnet 5 replaces the previous Sonnet 4.6 and, according to the company, comes very close to the significantly more expensive top-of-the-line model Opus 4.8. The improvement should be particularly noticeable when working independently with tools.

In Anthropic's model family, Sonnet sits between the compact Haiku and the more powerful Opus – it's the mid-range model for everyday use. With Sonnet 5, the company is now replacing Sonnet 4.6, released in February, which had occupied the mid-range position for a good four months. Anthropic describes the new version as the most agent-like Sonnet model to date: It's designed to plan, operate tools like browsers and terminals, and work autonomously at a level that, just a few months ago, was reserved for larger and more expensive models.

What distinguishes the new version

Anthropic sees the main progress in independent work. While the most significant leaps in agent capabilities were recently seen in the Opus class, Sonnet 5 is intended to close this gap: its performance is close to that of Opus 4.8, but at a lower price. Compared to its predecessor, Sonnet 4.6, the company cites clear improvements in logical reasoning, tool usage, programming, and general knowledge work.

This direction is not new. For many developers, the agent-based phase truly began with Sonnet models 3.5, 3.6, and 3.7, because these versions were the first to convincingly handle code and tools. Feedback from the early testing phase is now pointing in a similar direction: Sonnet 5 completes complex tasks that previous Sonnet models failed to address prematurely and verifies its own results without being explicitly prompted.

Security and vulnerabilities

Regarding security, Anthropic refers to its own assessments. According to these, Sonnet 5 exhibits less undesirable behavior overall than Sonnet 4.6 and is generally more secure in agent-based deployments. At the same time, the model's ability to perform cybersecurity tasks is significantly lower than that of current Opus models – a point deliberately kept low in view of potential misuse risks.

Availability and prices

Sonnet 5 is now available across all plans. It's the new standard model for the free and Pro plans, and is also available for the Max, Team, and Enterprise tiers. It's also available in Claude Code and via the Claude platform; developers can access it via API using the identifier claude-sonnet-5. There's an introductory price of $2 per million input tokens and $10 per million output tokens, valid until August 31, 2026. After that, the price will increase to $3 and $15, respectively. For comparison, Anthropic continues to charge $5 per million input tokens and $25 per million output tokens for Opus 4.8.

In parallel, Anthropic has increased rate limits across Chat, Cowork, Claude Code, and the Claude platform to account for higher token consumption with more intensive use. Adjustable workload levels allow users to choose the right balance between cost and accuracy by selecting between Sonnet 5 and Opus 4.8 – cheaper and faster with Sonnet 5, more precise and more expensive with Opus 4.8.

Where the Sonnet 5 stands in Anthropic's model range

Above the standard models, the Mythos class continues to be positioned. The latest Opus model, Opus 4.8, was released at the end of May, followed shortly thereafter by Claude Fable 5, the first publicly available model in the Mythos class. Fable 5 is the more secure variant, while the less restricted Mythos 5 remains reserved for a small circle of cyber defense providers and infrastructure vendors. Access was initially blocked at the behest of the US government; at the end of June, Mythos 5 was reopened to a select group of over 100 US institutions, and the company states that it is continuing to work on a broader rollout.

With Sonnet 5, Anthropic is pushing the mid-range tier closer to the performance previously reserved for the more expensive top-of-the-line models. In a market where OpenAI and Google are releasing new models at a rapid pace, this shift in the price-performance boundary is the real driving force – less a single benchmark than the question of how much capability will be available at what price in the future. (Image: Anthropic)