Fable 5.2 tops the benchmark but runs at batch speed

Claude Fable 5.2 leads the Artificial Analysis Intelligence Index with a score of 65, but reaching that number took an average of 45 minutes per task. Independent measurements show the model generated 275 million tokens across the nine evaluation suites, almost four times the median output of competing systems. At $8 per million input tokens and $40 per million output tokens, the benchmark run averaged $26.54 per task, paired with a time to first token of more than four minutes.

Those operational numbers define where this kind of reasoning architecture can actually live. A system that pauses for several minutes before streaming output is not built for human-in-the-loop interfaces or time-sensitive tool execution. Instead, this profile points to batch knowledge work: offline software verification, formal proof searches, or overnight analysis of extensive legal discovery documents, where waiting three quarters of an hour for a higher quality pass is an acceptable trade against human labor costs.

For teams building production agents, the calculation is straightforward. A top score on an intelligence index is useful signal, but latency and verbosity multiply quickly when an agent loops through external tools and API calls. Until extended reasoning becomes both faster and cheaper, the engineering challenge remains routing tasks so that these heavyweight models are called only when cheaper, faster models hit a wall.