How Claude Fable 5.1 took first place from Claude Fable 5 by a tenth of a point

On 1 September 2026, Index v4.8 placed Claude Fable 5.1 first, just above Claude Fable 5. The implementation record describes a score of 86.3 against 86.2, with the successor's point estimate at 93.1. Why would a much higher estimate produce a gap of only one tenth? The score was a conservative lower bound raised by a succession rule. These are the recorded v4.8 values, not the end-of-day result: v4.9 changed the interval later that same day.

The values recorded for v4.8

This source is a contemporaneous implementation record, not a retained upstream leaderboard snapshot. It is useful for explaining what the rule changed. It does not establish every launch-day benchmark value or the exact time a catalogue first listed the model.

What confidence actually measures

For measured v4 rows, confidence summarizes weighted support after identity, age and statistical adjustments. It is not a probability that the model deserves its rank. Before clipping or succession, the measured interval uses this formula:

Substituting the recorded confidence of 27.2 gives a nominal width of about 15.6 points. Half that width is about 7.8, so an estimate of 93.1 gives an unadjusted lower bound of about 85.3. The score scale clips endpoints at 100, so nominal width and displayed width need not match. The implementation record attributes the limited support to Arena Overall not yet listing the model; that source alone carries 60 points of the configured weight.

Why it is not simply #2

An unadjusted bound of 85.3 sits below the predecessor's 86.2. Sparse coverage can therefore reverse the order suggested by shared measurements. The Score v9 methodology record gives two examples of the newer model leading on the same evaluator's tests:

Index version 4.8 added a rule for exactly this case: a measured successor whose lower bound sat beneath its measured predecessor was floored one tick above it, and the row was labelled Measured · succession. The floor was withdrawn the moment the boards that had measured both models, weighted the way the score weights them, favoured the older release. The recorded result was the predecessor's 86.2 plus 0.1. The floor pass did not change the point estimate, which explains a score of 86.3 alongside a point estimate of 93.1. We wrote up the rule itself in a separate note.

What the gap does and does not say

The one-tick gap describes the ordering rule, not a measured capability difference. The two retained outside comparisons have different scales: 3.6 points on the Artificial Analysis index and 6.8 percentage points on its terminal test. Neither is interchangeable with a ModelCap point, nor establishes that the successor is better at every task. For a particular workload, the relevant test and configuration remain more informative than treating the floor as an effect size.

The v4.9 change later that day

The next implementation record changed more than the floor. Once a shared scored board supported succession, v4.9 could recompute the successor's interval using its measured predecessor's stronger support on boards the successor had not yet sat. The model still published its own confidence, with the inherited value recorded separately. The recorded preview moved Fable 5.1 from 86.3 to 90.7, with an interval of 90.7–95.5, while retaining first place.

That was a methodology change, not evidence that Arena had newly tested the model. It illustrates why a historical score needs a method version as well as a date. Under those versions, shared scored evidence favouring the predecessor withdrew the succession adjustment; new measurements could otherwise change both the estimate and support. Index v4.10 removed both rules on 4 September 2026, and since v4.11 Fable 5.1 has ranked on its own measured point estimate. This note makes no prediction of the next score or of when an evaluator will publish a result.

The method changelog in the methodology lists both versions; the two rows are Claude Fable 5.1 and Claude Fable 5.