Anthropic Releases Claude Sonnet 5: Focus on Autonomous Agents, Coding, and Hidden Costs

Anthropic has announced the release of Claude Sonnet 5, the latest iteration of its popular mid-tier language model. The new model is tailored for autonomous agentic scenarios, software development, interaction with external tools, and business process automation. It is available to all user tiers, including free Claude accounts.

Developers position Sonnet 5 as the most agent-ready solution in the lineup: the model handles action planning, terminal and browser operations, code self-verification, and autonomous execution without continuous human supervision much better. However, independent benchmarks indicate that the model tends to produce excessive intermediate reasoning, which can bring the total cost of completing a task close to the flagship Opus model.

Release Highlights and Pricing

Claude Sonnet 5 is already integrated into the Claude web interface, the Claude Code tool, the platform's API, as well as developer environments such as Cursor and GitHub Copilot. In most subscription plans, it is set as the primary default model.

For API access, the model has been assigned the identifier claude-sonnet-5. As part of an introductory promotional period lasting until the end of August 2026, discounted pricing is available: $2 per 1 million input tokens and $10 per 1 million output tokens. Once the promotion ends, the standard rate will be $3 and $15, respectively.

Anthropic's main objective with Sonnet 5 is to bridge the functional gap between the accessible Sonnet series and the heavyweight Opus tier as much as possible, shifting the focus from simple conversational queries to long, multi-stage workflows.

Why Agentic Metrics Matter More than Synthetic Benchmarks

The large language model industry is gradually moving away from text-based conversational assistants toward autonomous agents capable of modifying repositories, interacting with external software, and self-correcting errors when failures occur. When evaluating such models, the following parameters become critical:

• accuracy and correctness of the final solution;

• number of intermediate agent iterations and actions;

• total token consumption;

• real cost per completed project.

Performance and the Economics of Reasoning

In the Intelligence Index benchmark by Artificial Analysis, the model scored 53 points, coming close to top configurations like Opus 4.8 and GPT-5.5. However, the study revealed an important trend: in deep analysis mode, Sonnet 5 generates approximately 40% more output tokens compared to Sonnet 4.6, and in complex analytical scenarios, it executes up to three times as many agentic steps.

As a result, on a per-task basis, the total operational cost of Sonnet 5 after promotional rates end may reach the level of Opus 4.8, despite a formally lower price per token unit.

Changes in the Tokenization Mechanism

An additional factor contributing to higher costs is the redesigned tokenizer. Independent tests showed that an identical dataset is now converted into a larger volume of tokens: English texts see an increase of about 1.4x, Spanish about 1.33x, and Python code snippets around 1.28x (while the Chinese segment remains practically unchanged).

Anthropic representatives confirmed an increase in token density ranging from 1.0x to 1.35x depending on the data format, emphasizing that the introductory promo period is intended to offset this difference in team budgets.

Software Development Performance

In practical coding tasks, Sonnet 5 demonstrates clear progress:

• CursorBench: the model achieved a 57% success rate for edits compared to 49% in version 4.6 when working with multi-file codebases;

• CodeRabbit: the platform noted the model's improved ability to generate tests and design architectures, while review phrasing precision increased to 38–40% (up from 29%);

• Bug Detection: when reviewing third-party code, the percentage of identified defects dropped to 50–51% compared to 63% for Sonnet 4.6, while forcing deeper reasoning only doubles costs without a noticeable boost in review quality.

Model Comparison Across the Lineup

Safety

The model did not undergo specialized training in cybersecurity and lags behind specialized flagships like Mythos 5 in dedicated tasks. To prevent misuse, Sonnet 5 comes with integrated active defense systems designed to block harmful scenarios in real time.

Conclusion

Claude Sonnet 5 is a robust working tool for building agentic processes. When deploying the model, developers are advised to look beyond the listed price per million tokens and evaluate overall expenses across real-world workflows, taking into account longer reasoning chains and the characteristics of the new tokenizer.