SONNET 5 TEST: A good AI Allrounder for cheap by Anthropic?

2026-07-01 · Source: Discover AI · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Intermediate, medium

Summary

Anthropic released its new AI model, Sonnet 5, on June 30, 2026, positioning it as an "all-rounder." Benchmarks show Sonnet 5 performs at 43% without tools, trailing Opus 4.8's 49%, though it closes the gap with tool use. Its pricing is competitive, ranging from approximately \$2 per task for low performance to over \$20 for max performance, making it nearly on par with Opus 4.8 for agentic search. For agentic computer use, Sonnet 5 achieves 76-80% pass rates on OS World at 20-70 cents per task. However, on the newer OS World 2.0 benchmark (published June 28, 2026), Sonnet 5 is estimated to score below 20% binary accuracy, significantly underperforming Opus 4.8's 20-21%. Live testing on a free plan revealed a capped reasoning trace, an incorrect solution where the model invented rules, and a 21-minute correction time after user intervention due to a lack of "world model" understanding. The free plan also imposed a 5-hour usage limit.

Key takeaway

For AI Engineers evaluating new models for agentic computer use, Sonnet 5's estimated sub-20% performance on OS World 2.0 and its tendency to invent rules during complex reasoning tasks indicate it may not be suitable for critical applications. You should prioritize models demonstrating robust "world model" understanding and consistent accuracy on current, real-world benchmarks, even if free tiers offer initial access. Consider waiting for Opus 5 or Fable 5 for more reliable performance.

Key insights

Sonnet 5 offers competitive pricing but shows significant performance and reasoning limitations, especially on newer benchmarks and complex tasks.

Principles

• AI models can invent rules to justify incorrect outputs.
• "World model" understanding is crucial for complex tasks.
• Free tiers often have strict usage limits.

Method

The article describes a live testing setup involving a standard text input, observation of reasoning trace, and user correction for model errors.

In practice

• Evaluate models on real-world, up-to-date benchmarks.
• Test free tiers for unexpected limitations.
• Verify model outputs for invented rules or assumptions.

Topics

• Sonnet 5
• Claude Opus 4.8
• AI Benchmarking
• Agentic AI
• OS World 2.0
• Model Evaluation
• LLM Limitations

Best for: CTO, VP of Engineering/Data, AI Architect, AI Engineer, Machine Learning Engineer, Director of AI/ML

Related on AIssential

• I Tested Claude Sonnet 5 vs Opus 4.8 — Claude Sonnet 5 offers flagship-level agentic performance at 40% the cost, shifting focus from benchmarks to reliability in complex tasks.
• Is Sonnet 4.5 the best coding model in the world? — Leading coding AI models exhibit distinct reasoning styles, impacting performance and cost-efficiency in complex refactoring tasks.
• Claude Fable 5 - is this Mythos model worth the wait? — Claude Fable 5 offers superior benchmark performance and vision but struggles with prose, design, and conservative execution.
• Sonnet 5 review: I ran 64 generations to find out if it's worth it — Benchmarking frontier LLMs requires combining automated metrics with human "taste" to capture nuanced performance.
• [AINews] Sonnet 5 today, and Fable 5 tomorrow — Sonnet 5 improves agentic capabilities and coding, but its effective cost per task can exceed higher-tier models.
• Claude Sonnet 5: The Fable 5 at Home — Claude Sonnet 5 offers enhanced agentic capabilities, self-verification, and tool use, making it a cost-effective default for complex daily tasks.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Discover AI.