Notes
These notes explain how to interpret the board, why its ranking rules exist, and what changed after a release or an incident. Each article links to public methods, sources, or related analysis. Dated case studies describe the version in use at the time; current scores and availability are on the model pages.
A September 30 cutoff separates 131 measured-labelled rows from 148 modeled rows, records the top ten and source coverage, and identifies the evidence changes worth watching next month.
A retained September 30 snapshot shows how an evidence filter changes the price frontier. Exact token prices, workload arithmetic and an offline recipe make the trade-offs inspectable.
The public notebook reproduces all 279 current and 418 all-version positions from a retained input. A worked score explains the arithmetic and the upstream facts a rebuild cannot validate.
All 105 current coding observations at this cutoff come from Arena Coding. A refreshed shortlist separates normalized preference scores, tested efforts, overall ranks and workload prices.
In the retained September 30 cohort, 34 models have both general observations. Their median absolute normalized gap is 5.75 points; eight differ by more than ten. Settings and calibration limit the interpretation.
The best open-labelled model with at least 60% measured support sits 5.2 Index points behind the comparable API-only leader. Access categories, primary licences and self-hosting costs need separate checks.
The captured ledger report now has 12 settled first predictions, a 9.3-point mean absolute error and 75% interval coverage. Reproducible outcome rows expose both the misses and the proof limits.
A September 30 follow-up on Opus 5.5, GPT-6 Sol, Grok 4.7 and MiMo-V2.6-Pro. Broader evidence and product succession replace the old single-board story; missing launch-day inputs limit historical comparisons.
A worked guide to input and output pricing, the 3:1 blend, workload reversals and missing charges. The examples use invented rates so the method stays useful when live prices change.
A practical shortlist method: define the task, check identity and test settings, separate measurements from estimates, and record what public evidence cannot answer before choosing a route.
Under Index v4.8, Claude Fable 5.1 scored 86.3 against Fable 5's 86.2 despite a point estimate of 93.1. This explains that succession floor and the v4.9 change later that day; the one-tick gap was not a measured capability difference.
For about ninety minutes on 1 September 2026 the board hid Claude Fable 5, its measured #1, and opened Claude Fable 5.1 at #53 on a prior. Here is what the code did, why no alert fired, and the four rules that came out of it.
From 1 to 4 September 2026, Index versions 4.8 and 4.9 floored a measured successor one tick above its measured predecessor until the boards that scored both said otherwise. It moved DeepSeek V4 Pro from #102 to #20. The rule, the traps, what it did to the board, and why it was withdrawn.
Every rankable model opens on the board the cycle it is admitted. For open-weight releases the opening position comes from the publisher's own comparison table, read deterministically and quarantined. A worked example with Qwen3.8-27B, and the bug that nearly sent it to the corpus prior.
A guide to rank, score, support, intervals, evidence labels and movement under Index v4.14 and Score v10, and why measured and modeled rows need different readings.
Three of the coding and agent boards the Index scored have stopped publishing rows the board can use. The dormancy rule that retires them, the Terminal-Bench 4.0 change that broke the fetch, and why vendor-harness results are shown but never scored.
The first version of Score v9, shipped on 1 September 2026, admitted the Artificial Analysis runs of Terminal-Bench 2.1, Tau2-Bench Telecom and Humanity's Last Exam and capped each non-anchor evaluator at 30%. Neither survived the 4 September revision. What it did, and what replaced it.
Downloads, likes, prices, parameter counts, a publisher's own benchmark table, an LLM's opinion: each is available, each would make the board look more complete, and each is deliberately excluded from capability evidence. The reasons, and the one place a Hub error message forced us to be vaguer than we wanted.
A product should hold exactly one seat on the board however many ways it ships. Two rules get that right for API-plus-checkpoint twins. A third rule got it wrong for labs that release their only checkpoint in FP8, and quietly unseated GLM-5.3, DeepSeek V4 and gpt-oss until 1 September.
No human sits in the ranking loop and no scheduler lives in the source repository. A data plane discovers, admits, reads, scores, validates and publishes on a fixed cadence, keeps the last good board when a cycle fails, and reports its own health. The contract and the failures we have hit.