- INSTRUMENT
Leaderboard.
- FILED AS
Public benchmark boards, read from each board's own published table and reconciled to one model registry so a lab spelled two ways is counted once. Read the method.
- BOARDS
- 4
- RANKED ENTRIES
- 115
- LABS
- 14, on canonical identity
- SCRAPED
- 13 SEP 2026
The Consensus Board
On each board a family earns a percentile from where it sits among that board's canonical families: top of the board scores 100.0, bottom scores 0.0. The mean is those percentiles in the order above. Gap is the distance from a family's lowest board percentile to its highest. Tags come from the ten gaps on this table: the tightest quarter (10.2 points or less) reads UNANIMOUS, the widest quarter (37.5 or more) reads SPLIT. Today that is 3 unanimous, 3 split, and 4 that the boards have not ruled on. Composite koda-leaderboard-composite-v1.
| # | Model | Boards | The Spread0 = board floor, 100 = board top | Mean | Gappoints | LMArena | LiveBench | SWE-Bench Verified | Artificial Analysis |
|---|---|---|---|---|---|---|---|---|---|
| 1 | Claude Fable 5.1 | 3/4 | 96.0 | 12.0 | #4 | #1 | · | #1 | |
| 2 | Claude Fable 5 | 3/4 | 94.7 | 12.5 | #1 | #2 | · | #4 | |
| 3 | UNANIMOUS Muse Spark 1.3 | 3/4 | 84.3 | 9.7 | #7 | #4 | · | #5 | |
| 4 | Claude Opus 5 | 3/4 | 79.8 | 19.7 | #10 | #8 | · | #3 | |
| 5 | Muse Spark 1.2 | 2/4 | 69.6 | 28.8 | #5 | #14 | · | · | |
| 6 | SPLIT GPT-6 Astra | 3/4 | 68.3 | 79.8 | #25 | #3 | · | #2 | |
| 7 | UNANIMOUS Gemini 3.7 Flash | 2/4 | 66.5 | 5.0 | #12 | #10 | · | · | |
| 8 | UNANIMOUS Grok 4.6 | 2/4 | 66.5 | 8.8 | · | #12 | · | #8 | |
| 9 | SPLIT DeepSeek V4.1 Flash | 2/4 | 66.0 | 40.4 | · | #5 | · | #14 | |
| 10 | SPLIT GPT-5.6 Sol | 3/4 | 66.0 | 46.8 | #20 | #6 | · | #6 |
The Instruments
Every board read straight from source. The champion board leads at full width with price and throughput on screen; the other three sit beneath it. Each panel anchors its bars to a band computed from the rows it actually prints, so bar length is distance across that band, never from zero. A score column carries as many decimals as its own board publishes, capped at one, which is why an index ranked in whole points prints whole points beside a score ranked in tenths: no panel invents a resolution it was not given.
Who Owns The Board
Share of every top-20 canonical family slot across all 4 boards, by provider. One bar, 80 slots, the whole field at a glance. Percentages read as whole numbers on the bar and to one decimal on hover. 2 of those slots belong to labs with no validated palette colour and share the neutral segment: MBZUAI (1), Unknown (1).
Head to Head
Both models are lined up on every board they share. The bars are board percentiles, so two different scales can sit on one axis, and the delta is the distance between them: a measurement, not a verdict. Each side wears its own lab colour, or ink where both come from the same lab, and the exact score sits on hover. Boards where one is missing are skipped, not faked.
The Colour Contract
The one chromatic gesture on the podium and in the board cells. Where a lab wears the same hue as the medal, its dot renders neutral across the podium and the lab is named in words.
Ink, not hue. Three drawings on one 24 grid at one stroke weight, so the places separate without colour.
One validated hue per lab, on every dot, tick, bar and segment. 2 of the 80 slots belong to labs with no validated hue and render neutral rather than borrowing another lab's colour.
Ink and hairline. Eyebrows, category labels, source links, focus rings and row hover carry no hue of their own, so a colour inside this instrument always belongs to a lab.
Neutral outline, ink text, no fill. Age, head to head delta and rank movement are measurements, not verdicts. Nothing is past the 8 day window today.
Typographic, no hue. Ink for agreement, muted for disagreement, leading the row.
How This Is Counted
The stamp cannot precede its rows
The page prints DATA AS OF SEP 13, 2026 22:17 UTC, the newest scrape on it. The composite stamp is at or after every row it covers, so the Dateline square is filled.
Identity is a family, not a string
Each board contributes one row per family, its best ranked configuration. The raw board string stays beside it as LISTED AS, and where one board ranks two settings of the same product the rows carry the effort marker the registry recorded rather than repeating the name.
Age is recomputed in your browser
Every board carries its scrape stamp, so a page that stops rebuilding still ages on screen. The server value is the no-script fallback. Stale after 8 days.
Different boards measure different things
The consensus view averages family percentiles, not scores, and each panel measures its bars across a band computed from its own rows. Provider colours come from one validated palette; a lab with no entry renders neutral rather than borrowing another lab's colour. Sources: arena.ai, livebench.ai, swebench.com, artificialanalysis.ai.
Human preference rankings from blind A/B comparisons (7M+ votes)
Contamination-free benchmark with monthly-refreshed questions
Real-world software engineering task completion (Verified, 500 tasks)
Artificial Analysis Intelligence Index v4.3: aggregate of 10 evaluations