New Opus 5 half the price and more powerful than Fable 5? Tests explode the internet, users shocked

Anthropic officially released Opus 5, beating Fable 5 in multiple hardcore benchmarks at half the price. On Frontier-Bench, Opus 5 surpassed all competitors, more than doubling the performance of its predecessor Opus 4.8. In CursorBench at highest effort, it scored only 0.5% below Fable 5's peak. Users reported impressive results in coding, 3D modeling, and game development. Notable tests include a Rocket League clone, a 3D cell structure, a 360-degree rotating 3D cell structure, real-time wind tunnel simulation, and a Minecraft replica. Opus 5 also showed strong performance in CAD design and education. In a self-verification demonstration, Opus 5 rebuilt a 3D Boeing 747 model by extracting contour outlines and measuring 14 specific parameters against public data, while Fable 5 only checked screenshots by eye. Opus 5's price is $5 per million input tokens and $25 output, half of Fable 5. On ARC-AGI-3, Opus 5 scored 30.2% vs GPT-4.5 Sol's 7.8%. On Human's Last Exam without tools, Opus 5 got 56.3% vs Fable 5's 56.5%; with tools, Opus 5 achieved 64.7% vs Fable 5's 63.9%. Opus 5 also achieved a perfect score in IMO 2026 without external tools. Additionally, Anthropic reduced Claude Code's system prompt by over 80% without performance loss by leveraging Opus 5's improved judgment. Elon Musk posted a FrontierCode cost-performance chart showing only Grok 4.5 and Opus 5 on the Pareto frontier.