Anthropic shipped Claude Sonnet 5 on June 30, 2026, positioning it as the most capable mid-tier model for running autonomous coding agents at affordable prices. The model scores 63.2% on SWE-bench Pro, a benchmark that measures whether AI can resolve real GitHub issues by editing code across multiple files. That represents a 5.1-point jump from its predecessor, Claude Sonnet 4.6, and closes most of the gap to Anthropic's flagship Opus 4.8 at 69.2%.

What Makes Claude Sonnet 5 Different From Previous Versions?

Claude Sonnet 5 introduces several practical changes for developers integrating the model into their workflows. The most significant shift is that adaptive thinking, a reasoning feature that helps the model plan multi-step tasks, now runs by default. This means requests that previously required no reasoning budget will now spend tokens thinking through problems unless you explicitly disable the feature.

The model also ships with real-time cybersecurity safeguards built directly into the system, a first for Sonnet-tier models. This changes how developers must handle cases where the model refuses certain requests. Additionally, Claude Sonnet 5 launches across multiple platforms simultaneously: the Claude API, Amazon Bedrock, Google Cloud's Vertex platform, and Microsoft Foundry, with zero-data-retention support for organizations that require it.

How Does Claude Sonnet 5 Compare to Competing Models?

On the specific task of agentic coding, Claude Sonnet 5 outperforms both OpenAI's GPT-5.5 and Google's Gemini 3.1 Pro. Anthropic's internal evaluation shows Sonnet 5 at 63.2% on SWE-bench Pro, compared to roughly 58.6% for GPT-5.5 and 54.2% for Gemini 3.1 Pro. For teams building systems that edit files inside repositories, this represents a meaningful edge.

However, the competitive picture is not a clean sweep. Google's Gemini 3.1 Pro leads on web development tasks and multimodal workflows where visual context matters. The real strategic bet Anthropic is making is that "good enough to run unsupervised, at a price you can afford to run all day" beats "marginally smarter, but too expensive to loop." In other words, a slightly less capable model that costs less to operate continuously may deliver better real-world value.

On Terminal-Bench 2.1, which measures how well a model drives a command-line environment, Sonnet 5 jumps to 80.4% from Sonnet 4.6's 67.0%, a 13.4-point leap that speaks directly to the autonomous agent thesis. On computer-use tasks, Sonnet 5 scores 81.2%, and on knowledge-work evaluations, it actually edges past Opus 4.8 in one category, one of the few instances a mid-tier model has nudged ahead of its own flagship.

What Is the Pricing Strategy Behind This Launch?

Claude Sonnet 5 pricing is the fulcrum of the entire launch. Standard pricing is $3 per million input tokens and $15 per million output tokens, identical to Sonnet 4.6. However, Anthropic is running introductory pricing of $2 per million input tokens and $10 per million output tokens through August 31, 2026, after which the model reverts to standard rates.

The value proposition is straightforward: you get materially more capability at the same sticker price you were already paying for Sonnet 4.6. For teams running thousands of autonomous agent loops daily, where output tokens dominate the bill, the difference between frontier-class pricing and mid-tier pricing compounds quickly. According to TechCrunch, Sonnet 5 undercuts both OpenAI's GPT-5.5 and Google's Gemini 3.1 Pro on price, while Google's Gemini 3.5 Flash remains cheaper still. This places Sonnet 5 in a deliberate sweet spot: cheaper than frontier-class competition, more capable than the budget tier.

Steps to Evaluate Claude Sonnet 5 for Your Coding Workflows

• Benchmark Against Your Codebase: Test Sonnet 5 on real GitHub issues from your own repositories to see how well it handles your specific coding patterns and file structures, rather than relying solely on published benchmarks.
• Calculate Your Token Costs: Estimate how many input and output tokens your typical agent loops consume, then multiply by the introductory $2/$10 rates through August 31 to understand your cost baseline before standard pricing kicks in.
• Test Autonomous Behavior: Run multi-step tasks that require planning, tool use like terminals and browsers, and file editing across multiple files to verify the model's agentic capabilities match your use case.
• Review Security Safeguards: Examine how the new real-time cybersecurity features affect your workflows, particularly how the model handles refusals and whether you need to adjust your error handling logic.

The benchmark numbers come with an important caveat: these are Anthropic's own measurements, published alongside the model on the company's Transparency Hub. Vendor-run benchmarks are useful directional signals, not neutral referees. Independent aggregators such as LM Council typically take a few weeks to publish reproductions. The relative gains, where Sonnet 5 clearly beats Sonnet 4.6 and trails Opus 4.8 by single digits, are the safest read. The absolute figures may move once third parties re-run them.

Claude Sonnet 5 represents a deliberate shift in how Anthropic is positioning its product line. Rather than competing on raw capability alone, the company is betting that developers will choose a model that balances performance with affordability, especially when running agents autonomously at scale. The introductory pricing through August 31 creates a window for teams to evaluate whether the model's capabilities justify adoption before standard rates apply.