Anthropic Releases Claude Sonnet 5: Performance Close to Opus 4.8, Lower Price, Agent-Focused

Anthropic has officially launched Claude Sonnet 5, positioning it as the "most agentic Sonnet model to date." The model shows significant improvements over its predecessor Sonnet 4.6 in reasoning, coding, tool use, and knowledge work, with benchmark scores approaching the flagship Opus 4.8, while its API price is only 60% of Opus 4.8 (with an even lower introductory price). Sonnet 5 is now available across all platforms, becoming the default model for Claude Free, Pro, Max, Team, and Enterprise users, and supports a 1M token context window.

Performance

Sonnet 5 surpasses Sonnet 4.6 on several key benchmarks and approaches Opus 4.8:

Agentic Coding (SWE-bench Pro): Sonnet 5 scores 63.2%, up from Sonnet 4.6's 58.1%, below Opus 4.8's 69.2%.
Multidisciplinary Reasoning (Humanity's Last Exam): Without tools, Sonnet 5 scores 43.2% (Sonnet 4.6: 34.6%, Opus 4.8: 49.8%); with tools, it rises to 57.4%, close to Opus 4.8.
Computer Use (OSWorld-Verified): Sonnet 5 scores 81.2%, Sonnet 4.6: 78.5%, Opus 4.8: 83.4%.
Agentic Search (BrowseComp): At high/xhigh/max levels, Sonnet 5 performs close to Opus 4.8.
CursorBench 3.1: Sonnet 5 scores 57%, Sonnet 4.6: 49%, close to Opus 4.8 high.

Third-party benchmark Artificial Analysis Intelligence shows Sonnet 5 max scores 53, on par with GPT-5.5 high, below Opus 4.8 high and GPT-5.5 xhigh.

Pricing and Cost

Standard pricing for Sonnet 5 is $3 per million input tokens and $15 per million output tokens; until August 31, 2026, the promotional price is $2 input and $10 output, approximately 40% of Opus 4.8 ($5 input, $25 output).

Actual usage cost varies by task. For example, in a comparison test building a single HTML login page:

Sonnet 5: 20.9k input tokens, 14.2k output tokens, total cost $3.36, time 2 min 11 sec.
Opus 4.8: 96.3k input tokens, 73.8k output tokens, total cost $20.66, time 20 min 15 sec.

However, on a Cost per Intelligence Index Task basis, Sonnet 5 max costs $2.29 per task, higher than Opus 4.8 max's $1.80, indicating actual cost is influenced by output volume, reasoning depth, etc.

New Features and Notes

Adaptive Thinking: Replaces extended thinking mode, defaults to medium effort, automatically adjusts based on task.
Tokenizer Update: Same text maps to more tokens (increase factor ~1.0-1.35x); Anthropic states the promotional price aims to keep migration costs roughly equal.
Rate Limit Increase: To accommodate higher token consumption from increased effort modes, Anthropic has raised rate limits for Chat, Cowork, Claude Code, and the platform.
Safety Evaluation: Sonnet 5 outperforms Sonnet 4.6 in rejecting malicious requests, resisting prompt injection, hallucination rate, and sycophancy, but has a slightly higher inappropriate behavior rate than Opus 4.8 and Mythos Preview.

Availability

Sonnet 5 is available across all platforms, including the native Claude platform, AWS, Google Cloud, Microsoft Foundry, etc. Claude Free and Pro users automatically switch to Sonnet 5 as the default model; Max, Team, and Enterprise users can also use it. Developers can access it via Claude Code and the Claude Platform API.

Industry Feedback

Early access partners unanimously report that Sonnet 5 is more autonomous and agentic than its predecessor, capable of completing complex tasks, with an attractive price. Cursor has announced support for Sonnet 5.

Summary

The release of Sonnet 5 marks the migration of agentic capabilities from flagship models to mid-range models. For cost-sensitive teams that need stable execution of multi-step tasks, Sonnet 5 becomes the new default; for tasks requiring high accuracy, Opus 4.8 remains the top choice.