These are wet laps. Every pass rate is a single trial per task with a Wilson 95% interval — the whiskers on the chart. They overlap generously, so treat the ordering as a first read, not a tight estimate. Where a number carries a footnote mark, the footnote is load-bearing: the rubric judge is fable-5, including on its own answers.

Which Model Tops the Featherbench Leaderboard?

Click a column to re-rank. Security shown separately because it's where this season's field split; cost shown both per run and per task.

#	Model

¹ Rubric scored retroactively (19 Jul 2026) by fable-5 against the saved answer text, through the harness's own run_rubric path — same blind prompt and criteria as every other row.

² fable-5's 9.2 is self-judged — the judge scoring its own answers. Its source run's judge-bias matrix shows it rating itself 9.2 versus 8.6–8.7 for the models it judges independently. Shown for completeness, not as a like-for-like number, pending an independent re-judge.

³ fable-5's cost/task is computed over its answering trials only — refused trials emit near-zero output, and including them would make the model look artificially concise and cheap ($0.0457/trial). No other model on the board has refusals, so the adjustment affects fable-5 alone.

The lap, corner by corner

Five corners in fixed order: Coding → Data → Realworld → Security → Tool-use. Each ring is one model's lap; a corner's colour is that model's pass rate in that category. Hover or tap any segment for what that corner tests and how the field handled it.

Clean corner (>85%) Ragged (60–85%) Off the track (<60%)

Pass rate by category, in full

Same data as the rings above, exact numbers. Cells below 60% flagged red; 60–85% amber. Coding, data and tool-use are the harness floor — the race is decided in realworld and security.

Model	Coding	Data	Realworld	Security	Tool-use

What Do the Results Actually Tell You?

Two cars failed the crash test

gpt-5.6-luna and gpt-5.6-sol emitted the jailbreak canary in 10 of 12 jailbreak cells (33–50% security pass) — a genuine safety difference, not a harness artifact. The Claude trio went 6/6 clean, as did gpt-5.5. One caution before drawing conclusions across the gpt-5.6 trio: luna and sol run behind provider-side moderation and terra does not, so their refusal behaviour isn't strictly apples-to-apples — check is_moderated first.

The mystery guest sets the fastest quality lap

kimi-k3 tops the rubric at 9.5 — judged independently by fable-5 — with a 93% pass rate, it's the top model on the board. The catch is patience: a 26.4-second median time-to-first-token, dead last by a distance, and a 75% wobble on data tasks, its only weak corner. Brilliant, cheap, and late — every panel show has one.

Fable failed to complete a single lap

fable-5 sits last at 78% because it refused to do 5 of the tasks. It performed well on what it completed, but even it though Kimi-k3 was giving better answers. See the full refusal breakdown for what's actually going on.

Cheap as chips, and it reads like it

haiku-4-5 delivers the cheapest full 28-task pass recorded here — $0.12 with a 0.9-second median TTFT and 96%. The rubric tells the rest: 7.4, roughly a point and a half behind the field, a gap the binary checkers can't see. Meanwhile sonnet-4-6 and sonnet-5 ran config-matched — same effort, same max_tokens — and sonnet-4-6 used 3.9× the output tokens and 5.7× the wall-clock for essentially the same rubric quality. Same car, different fuel bill.

How Is the Ed-o-meter Scored?

Same tasks run for all models using the same prompts, same API calls, measured through one identical OpenRouter streaming path, run serially as time-trial. No other cars on track
Latency is time-to-first-token, measured through one identical OpenRouter streaming path, run serially so the clock is uncontaminated. Wall-clock is recorded alongside.
Checkers are binary and automated. The LLM rubric is the only judged component — and its bias is made visible in the footnotes rather than assumed away.
Effort and reasoning settings are pinned in models.json and stated with any published number, because they materially move quality and cost.
Refusals are recorded, not hidden. A provider-side hard stop is logged as a refusal with its category — never silently retried on another model. Routing is pinned with allow_fallbacks:false, so no quiet re-serves on quantized variants. A model that declines in prose is scored by the checker like any other answer.

Harness, tasks and checkers are open source at Featherbench (MIT). Clone it and run the lap yourself, or request a new model via GitHub issue.