Claude Fable 5.2 scores 65 on the Artificial Analysis Intelligence Index v5.0.1 — first of 78 models, five points clear of second-placed Gemini 4 Carbon. OrcaRouter's November 13 article takes apart what that score costs: on the test harness, 267 seconds to first token, 46.8 minutes per task and $26.54 per index run — while the same model on the same company's live gateway shows a median first token of 8.09 seconds. Both numbers are correct; they measure two very different ways of using the model.
The index is nine evaluations: AA-Briefcase, AutomationBench-AA, Terminal-Bench 4.0, FrontierCode-AA 2.0, AA-Omniscience, GDP.pdf, HLE-Diamond, CritPt v2 Hard Subset and Terminal-Bench-Science — three terminal and automation tasks, two document-grounded reasoning tasks over long inputs, two hard science and mathematics evaluations, one frontier code evaluation, and no web-agent harness, multi-turn support simulation or vulnerability detection. On that set, Fable 5.2 takes 267 seconds to first token (76th of 77 models with latency data) and 46.8 minutes per task. Output, once it starts, is not slow: 97 tokens per second, 35% faster than Carbon's 72, but that speed accounts for almost none of the elapsed time. The money is more extreme still: at $8.00 per million input and $40.00 per million output tokens, one index run costs $26.54 — the most expensive of 77 models, against $5.48 for Carbon on the same nine evaluations. That is 4.8 times the spend for five index points, or $0.41 of benchmark spend per index point against Carbon's $0.09. OrcaRouter calls it the single most useful ratio in the comparison: at this price, the model is a tier, not a default.
The same model looks entirely different in OrcaRouter's live gateway traffic: across 35.6 million tokens over seven days, first token comes at a median of 8.09 seconds and a p95 of 10.00 seconds. The gap is not in the model but in what is being measured: the harness lets the model think at maximum effort on hard problems such as HLE-Diamond, while the gateway measures the median request at whatever effort the caller set. OrcaRouter puts it plainly — run at max effort and you will reproduce AA's number; run at low effort and you will see something near the gateway's.
AA reports 275 million output tokens for the run with verbosity at 0.97%, against Carbon's 205 million tokens and 2.69%. The percentage and the count point in opposite directions, so the percentage should not be quoted as evidence of concision. The ratio worth remembering is the one from the previous section: $0.41 of spend per index point against Carbon's $0.09.
The index does not measure reliability; live traffic exposes it: OrcaRouter saw a 10.8% non-2xx rate across all upstream providers over seven days, against 39.7% for Carbon (already noted in our Carbon article). A benchmark harness absorbs that with retries; your application will not unless you build it. Note that 10.8% is an all-provider aggregate, not an error rate for Fable 5.2 alone.
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 also 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.
OrcaRouter's advice is to route by difficulty: 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 — hand the majority to a cheaper model. If your task is interactive, the harness numbers are not your numbers: design against the gateway p50 of 8.09 seconds, and the effort level is the variable that decides whether you land near eight seconds or near four minutes. One more discipline on versions: only models re-scored on v5.0.1 belong in the same table; mixing in v4.3 or v4.3.2 scores invalidates the whole chart.
Reference source: this article is an original rewrite produced by this site; facts and figures were checked against OrcaRouter — Claude Fable 5.2 benchmark (Nov 13, 2026)
← Back to the Claude Fable 5.2 topic page