Two rows on the Claude Fable 5.1 card both measure intelligence, both carry `artificialanalysis.ai` as their source, and they differ by 12.6 points: `AA Intelligence` reads 53.4 and `Artificial Analysis Intelligence Index` reads 66 [OURS]. Nothing on the card explains the difference. The fable 5.1 price query lands on this table. The card is served by OrcaRouter.
The two rows, and why a buyer should care
A card that prints two numbers for one concept is not broken. It is usually printing two versions of an index that the evaluator revised, kept both names for, and never deprecated on the vendor side.
But a buyer comparing cards sees only one of them at a time, and the choice changes the picture materially. At 53.4, Fable 5.1 sits just above Claude Opus 4.6’s class; at 66 it is in a different tier entirely.
The practical rule is that the lower number is the conservative one and the one to quote when comparing across cards, because it is the row that appears on more cards in this catalogue. The 66 is useful as a within-card trend, not as a cross-card rank.
The generation delta, on the rows both cards share
Fable 5.1 publishes ten benchmark rows; Fable 5 publishes nineteen. Nine rows overlap, which makes a clean generation comparison possible:
| Row | Fable 5 | Fable 5.1 | Change |
| AA Coding | 76.5 | 81.6 | +5.1 |
| AA Intelligence | 49.6 | 53.4 | +3.8 |
| GPQA Diamond | 92.6 | 93.7 | +1.1 |
| Humanity’s Last Exam | 55.5 | 59.1 | +3.6 |
| Long-Context Recall | 82.3 | 85.3 | +3.0 |
| SciCode | 61.0 | 63.1 | +2.1 |
| terminalbench_v2_1 | 84.6 | 91.4 | +6.8 |
| terminalbench_v4_0 | 42.4 | 52.0 | +9.6 |
| tau_banking | 38.1 | 47.2 | +9.1 |
Every row moved up, which is not guaranteed between two releases of the same family. Two of them moved a lot.
terminalbench_v4_0 from 42.4 to 52.0 is the largest gain, on the hardest generation of that benchmark. For context, most cards in this catalogue score in the single digits to low thirties on v4.0, and 52.0 is the highest reading in this batch [OURS].
tau_banking from 38.1 to 47.2 is a 9.1-point move on a tool-use benchmark where scores are generally low. Both cards are far from saturated there, so the gain is measured where the instrument still discriminates.
The rows that barely moved are the ones already near their ceiling: GPQA up 1.1 and SciCode up 2.1 describe a model that was already good at those and is now slightly better.
What did not improve
Three things on the card got worse or stayed flat, and they are easier to miss than the benchmark rows.
The error rate. Fable 5 ran at 0.61 percent over its seven-day window; Fable 5.1 reports 41.43 percent [OURS]. That is not a model regression — it is a route that is not yet steady — but it is the number that decides whether the improvements above are usable.
The published row count. The older card carries nineteen rows, including seven vendor-published hard tasks. The newer card carries ten, all from the evaluator. The vendor’s own hard-task rows — OSWorld, SWE-Bench Pro, GDPval.pdf, FrontierCode, AutomationBench, Legal Agent Benchmark — are gone.
That is a real loss of information. A reader of Fable 5.1 cannot see how it does on the tasks Anthropic itself considers hard, only how it does on the benchmarks an evaluator chose.
Output speed. 74.3 tokens per second on the older card against 83.5 on the newer — an improvement, but small enough that it does not change what the model is for.
How to read a card with two intelligence numbers
Four rules, in order.
Quote the row that appears elsewhere. If every other card in your comparison prints `AA Intelligence`, use 53.4 and set the 66 aside.
Prefer rows both generations carry. The nine overlapping rows here are the only place the delta is measurable; a row only one card has cannot show a generation change.
Check what disappeared. A card that drops seven vendor rows is not hiding failures — it is showing less. Treat the absence as missing data, not as a clean sheet.
Read telemetry before benchmarks. A 12.6-point spread between two index rows is a curiosity next to a 41 percent error rate, which decides whether any of it can be used.
The takeaway
Claude Fable 5.1 prints two intelligence scores from the same evaluator — 53.4 and 66 — and gains on all nine rows it shares with Fable 5, including +9.6 on terminalbench_v4_0 and +6.8 on terminalbench_v2_1 [OURS].
The gains are real and the card shows less than its predecessor did: nineteen rows down to ten, and a seven-day error rate that went from 0.61 percent to 41.43. The benchmarks say the model improved; the telemetry says the route has not caught up.