Where Claude Opus 5 Fits in Your Model Rotation

What It Covers

Anthropic releases Claude Opus 5, positioned between Opus 4.8 and Fable 5 in capability and cost, scoring state-of-the-art results on ARC-AGI 3 at 30.2% and OS World 2.0 at 70.6%, while generating mixed real-world feedback around personality, reliability, and workflow compatibility.

Key Questions Answered

• •Effort settings calibration: Opus 5 performance peaks at "extra high" settings, not "max." On FrontierBench and the Artificial Analysis coding index, max settings cause the model to over-verify, spin on simple problems, or stray beyond task scope. Set effort to extra high for best performance-to-cost ratio, saving 36% versus Fable 5.
• •Context engineering overhaul required: Anthropic stripped 80% of Opus 5's built-in system prompt, finding over-constrained instructions conflicted with user workflows. Existing skills libraries built for older Claude models will break. Rewrite prompts using progressive context disclosure, descriptive language over examples, and minimal front-loading. Use the new "Claude Doctor" command to automate cleanup.
• •Enterprise model rotation logic: Opus 5 inherits Opus 4.8 pricing at $5 per million input tokens and $25 per million output tokens, costs roughly half of Fable 5 per task on CursorBench, and crucially lacks Fable's data retention policy — making it viable for sensitive enterprise data where Fable is a non-starter.
• •Benchmark-to-practice gap is widening: Opus 5 leads Fable 5 by 146 ELO on the AA Briefcase long-horizon knowledge work benchmark, yet multiple practitioners found it stops tasks early, argues with instructions, and breaks existing workflows. Treat benchmark rankings as directional signals only; run task-specific blind tests before committing to model rotation changes.
• •ARC-AGI 3 score demands scrutiny: Opus 5 scored 30.2% on ARC-AGI 3, demolishing GPT-5.6 Soul's 7.8%. However, researchers note Anthropic trained the model on reinforcement learning environments resembling ARC-AGI puzzles using human contractor reasoning traces. Treat this score as potentially reflecting training overlap rather than pure out-of-distribution generalization capability.