Anthropic Unveils Claude Opus 5: Elite Performance at Half the Cost

Anthropic has officially launched Claude Opus 5, a new flagship model designed to disrupt the high-end LLM market by offering near-parity with the premium Claude Fable 5 at a significantly lower price point. This strategic move aims to neutralize pricing pressure from competitors like GPT-5.6 Sol while providing developers with a more efficient tool for coding and complex knowledge work.

Disrupting the Price-Performance Frontier

The most striking feature of Claude Opus 5 is its aggressive pricing structure. While the top-tier Claude Fable 5 costs $10 per million input tokens and $50 per million output tokens, Opus 5 slashes these rates in half, priced at $5 per million input and $25 per million output tokens.

To address latency requirements, Anthropic has also introduced a "Fast Mode," which increases processing speed by 2.5x, though it comes at the cost of doubling the token price. This tiered pricing, combined with a massive 1 million-token context window, positions Opus 5 as the new default model for Claude Max and the most capable option for Claude Pro users.

Benchmark Supremacy in Coding and Reasoning

Opus 5 is not merely a budget alternative; it is a performance powerhouse in specific domains. In agentic terminal coding via the Frontier-Bench v0.1, Opus 5 achieved a score of 43.3%, significantly outperforming both Fable 5 (33.7%) and GPT-5.6 Sol (34.4%). It also demonstrated dominance in knowledge work, leading the GDPval-AA v2 benchmark with an Elo score of 1,861.

Perhaps the most staggering statistic comes from the ARC-AGI-3 benchmark, which measures novel problem-solving. Opus 5 scored 30.2%, nearly quadrupling the 7.8% achieved by GPT-5.6 Sol. This suggests a massive leap in the model's ability to handle tasks that fall outside of its training data patterns.

Intelligent Effort Scaling and Efficiency

Anthropic has introduced a sophisticated "effort" system, allowing users to toggle between five settings: low, medium, high, xhigh, and max. This allows for a granular trade-off between token consumption and reasoning depth.

While Anthropic recommends "low" and "medium" settings for general tasks to optimize costs, they advise using "xhigh" for complex agentic coding. Interestingly, the company noted a diminishing return at the "max" setting; on both Frontier-Bench v0.1 and the Artificial Analysis Coding Agent Index, Opus 5 actually saw a slight performance dip at maximum effort despite the higher cost, suggesting that the "xhigh" setting may be the current sweet spot for efficiency.

Improved Safety and Autonomy

In a direct response to user feedback, Anthropic has fine-tuned the safety classifiers for Opus 5. The model's cyber filters trigger approximately 85% less often than those on Fable 5, addressing previous criticisms regarding over-zealous blocking of legitimate research. While it still prevents exploit generation and binary-based scanning, it is far more permissive for source code vulnerability research.

Furthermore, Opus 5 shows enhanced self-correction capabilities. In real-world testing, the model successfully built a 3D model in FreeCAD by writing its own computer vision pipeline to interpret geometry—a task that stumped all other competing models.

Key Takeaways

  • Drastic Cost Reduction: Opus 5 offers flagship-level performance at 50% of the token cost of Claude Fable 5.
  • Reasoning Breakthrough: The model showed a massive 4x lead over competitors in the ARC-AGI-3 benchmark for novel problem-solving.
  • Optimized Developer Workflow: With improved safety classifiers and specialized "effort" settings, Opus 5 provides a more fluid experience for agentic coding and research.