Claude Fable 5.2 Benchmark: What the 65 Costs in Wall Clock

Claude Fable 5.2 scores 65 on Artificial Analysis Intelligence Index v5.0.1 — first of 78 models, five points clear of Gemini 4 Carbon. Those five points come with 267 seconds to first token, 46.8 minutes of wall clock per task and $26.54 of spend per index run, every one of which is the practical shape of the score. This is what the harness measured and what to do with it. The Claude Fable 5.2 API page has the endpoints and prices; this piece is about the benchmark.

The nine evaluations

AA-Briefcase v1.1, AutomationBench-AA, Terminal-Bench 4.0, FrontierCode-AA 2.0, AA-Omniscience v1.1, GDP.pdf, HLE-Diamond, CritPt v2 Hard Subset, Terminal-Bench-Science 0.1. Three terminal and automation tasks, two document-grounded reasoning tasks over long inputs, two hard science and mathematics evaluations, one frontier code evaluation. No web-agent harness, no multi-turn support simulation, no vulnerability detection.

The score and its scale

65, rank 1 of 78, on v5.0.1. Only models re-scored on v5.0.1 belong in the same table; v4.3 and v4.3.2 scores are a different evaluation set and any chart that mixes them is invalid. Within the current cohort the ordering at the top is Fable 5.2 at 65 and Gemini 4 Carbon at 60, with the rest of the field further back or not yet re-scored.

A five-point gap on a nine-evaluation index is real but narrow. The useful reading is not 'five percent better' but 'a modest edge on a hard suite', and the reason to pay for it is variance rather than raw capability: on tasks where a wrong answer invalidates the whole run, a small edge in success probability compounds across steps.

The wall clock

267 seconds to first token, rank 76 of 77 models with latency data. 46.8 minutes total per task. 97 tokens per second of output once it starts, which is 35 percent faster than Carbon's 72 and accounts for almost none of the elapsed time.

The distribution is the important part. Through OrcaRouter's live traffic the same model reports a p50 of 8.09 seconds to first token and a p95 of 10.00 seconds over 35.6 million tokens in seven days. The harness and the gateway are measuring different populations: one lets the model think at max effort on HLE-Diamond, the other is the median request at whatever effort the caller set. If you run Fable 5.2 at max effort you will reproduce AA's number; at low effort you will see something near the gateway's.

The money

$26.54 per index run at $8.00 per million input and $40.00 per million output — rank 77 of 77, the most expensive on the board. Gemini 4 Carbon costs $5.48 for the same nine evaluations. That is 4.8 times the spend for five index points, and it is the single most useful ratio in this comparison.

AA reports 275 million output tokens for the run with verbosity at 0.97 percent. Note that the percentage and the count point in opposite directions relative to Carbon's 2.69 percent and 205 million tokens, which is why the percentage should not be quoted as a concision claim.

The ratio that makes the comparison usable is $0.41 of benchmark spend per index point against Carbon's $0.09. Whatever you conclude from it, the arithmetic forces a routing decision rather than a single-model decision: at this price the model is a tier, not a default.

What the score does not tell you

There is no cyber evaluation in the set, so nothing here speaks to vulnerability recall or patch quality — relevant if that is why you are looking at a frontier model. There is no long-running agent evaluation with tool calls across turns, which is the workload Fable 5.2 is actually marketed for; the terminal tasks approximate it but do not reproduce it.

There is also no reliability measurement. Live traffic on OrcaRouter showed a 10.8 percent non-2xx rate across all upstream providers over seven days, against 39.7 percent for Carbon. A benchmark harness absorbs that with retries; your application will not unless you build it.

Reading it as a decision

If your task looks like the nine evaluations and you can run it asynchronously, 65 at $26.54 per run is defensible for the hard cases and wasteful for the easy ones — route by difficulty and use a cheaper model for the majority. If your task is interactive, the harness numbers are not your numbers; the gateway p50 of 8.09 seconds is the one to design against, and the effort level is the variable that decides whether you land near eight seconds or near four minutes.

For the head-to-head, Claude Fable 5.2 vs Claude Fable 5.1 covers the previous generation at a higher price, and Gemini 4 Carbon vs Claude Fable 5.1 covers the cross-vendor pair. The LLM models list has the rest of the v5.0.1 cohort with ids and prices; re-check the index version on any chart you reuse, because a v4.3 score in the same table will invalidate the comparison.