Anthropic launches Claude Sonnet 5 for more affordable AI agents

Anthropic released Claude Sonnet 5 on 30 June 2026, a model that combines stronger agentic capabilities with a lower price point than the heavier Opus model. The latest Sonnet model is explicitly aimed at those who want to build and run autonomous AI agents without the bill that comes with an Opus or GPT-5.5 subscription.

Sonnet 5 launches with an introductory price of $2 per million input tokens and $10 per million output tokens, valid through 31 August 2026. After that, rates of $3 and $15 respectively will apply. For comparison: Opus 4.8 costs $5 per million input tokens and $25 per million output tokens. With prompt caching, Anthropic says savings of up to 90% are possible; batch processing delivers up to a 50% discount.

The model is immediately available via the Anthropic API, on Claude.ai for free and paid subscriptions, and through Amazon Bedrock and Google Cloud Vertex AI. Developers can also use the model in Claude Code and in the Kiro IDE.

What Sonnet 5 adds compared to earlier Sonnet versions

Anthropic describes Sonnet 5 as the most agentic Sonnet model to date. The model can plan multiple steps, call tools such as browsers and terminals, and independently carry out complex tasks that earlier Sonnet generations could not handle.

On the agentic coding benchmark, Sonnet 5 scores 63.2%, compared with 58.1% for Sonnet 4.6 and 69.2% for Opus 4.8. On the GDPval-AA v2 benchmark for knowledge work, the model achieves 1,618 points, versus 1,395 for Sonnet 4.5. In absolute terms, Opus 4.8 therefore remains the stronger model, but Sonnet 5 narrows the gap while remaining less expensive.

The context window is 1 million tokens; the maximum output per call is 128,000 tokens. Anthropic has also updated the tokenizer, which improves text processing. A side effect is that the same input may produce 1.0 to 1.35 times more tokens than in previous versions, something to account for in cost estimates.

Safety profile changes significantly compared to Sonnet 4.6

Anthropic highlights improved safety as a distinguishing feature of Sonnet 5. Specifically: the success rate of prompt injection attacks dropped from around 50% on Sonnet 4.6 to less than 1% on Sonnet 5, and effectively to 0% when the built-in safeguards are active.

The model also shows fewer hallucinations and less sycophantic behaviour than its predecessor. Anthropic has built in real-time cybersecurity measures that detect and block dangerous use, comparable to what was already available in Opus 4.7 and 4.8. The capacity for dangerous cyber tasks is considerably lower in Sonnet 5 than in Opus models, although Anthropic emphasises that the model has not been specifically optimised for cybersecurity applications.

For companies running agentic workflows, where a model independently executes multiple steps and calls external tools, the reduced vulnerability to prompt injection is a relevant consideration. Malicious actors can use that attack vector to try to steer an agent towards unintended actions.

Availability and early users

Sonnet 5 becomes the default model for free and Pro subscriptions on Claude.ai, available via web, iOS and Android. Max, Team and Enterprise users also have access, as do users of Claude Code and AWS's Kiro platform, in the US-East-1 and Europe (Frankfurt) regions.

Anthropic names a number of early users, including co-founders of Lovable and Eve, engineers at ClickHouse, and several individual developers. No concrete performance figures or case studies from those collaborations were published at launch.

The model's knowledge cutoff is January 2026, a practical consideration for anyone deploying it on tasks where up-to-date information is relevant.

What this means for those building AI agents

The launch of Sonnet 5 fits into a broader trend in which frontier labs continue to upgrade their mid-tier models, shrinking the performance gap with their heaviest offerings. For builders looking to run cost-efficient agentic pipelines, a model that approaches Opus-level performance at a lower price represents a meaningful difference in feasibility.

A full comparison with competitors such as GPT-5.5 and Gemini Pro is difficult to make at this point, as benchmark figures and pricing structures for those models have been published on their own terms and may change over time. What is concrete are Anthropic's own benchmarks and the published prices for Sonnet 5 and Opus 4.8.

For those currently working with Sonnet 4.6, the improved resistance to prompt injection is the most directly applicable improvement, especially in production environments where agents have access to sensitive systems or data.