I Read Six AI Leaderboards on One Day. They Named Six Different Number Ones.

TL;DR: On 2026-08-15 I fetched seven public AI leaderboards and recorded the number-one entry on each. Six were readable, and they named six different entries at rank 1, spanning three different base models: Claude Fable 5, Claude Opus 5 and Muse Spark 1.1. Three of the seven publish an error term, and on all three the gap between rank 1 and rank 2 falls inside that published interval. There is no model that is "number one right now," because the boards are not measuring one quantity.

Every launch post this year says state of the art. I wanted to know how many models that is true of at once, so I picked a day and read the boards.

This is a snapshot. The date is 2026-08-15, every figure below comes from a fetch I ran that day, and some will be stale by the time you read this. The method is at the bottom.

What did the seven boards actually say?

Leaderboard	Rank 1 on 2026-08-15	Score	Error term published?
Arena text (default style-control view)	claude-fable-5	1506 Elo, shown as ±5	Yes, plus a rank spread of 1 to 4
Artificial Analysis Intelligence Index	Claude Opus 5 (max)	63	No
ARC Prize, ARC-AGI-3 default tab	Claude Opus 5 (High)	30.16%	No
Humanity's Last Exam, live board	Fable 5	52.72%	No
GDPval-AA v2	Claude Opus 5 (Adaptive Reasoning, Max Effort)	1849 Elo, -22 / +22	Yes
SWE-bench Pro, public set	Muse Spark 1.1	61.50, ±3.10	Yes
OSWorld-Verified	Could not read, see below	n/a	n/a

Six readable boards, six different strings at rank 1, three base models underneath them: four of the six winners are a reasoning-effort configuration of Claude Opus 5 or Claude Fable 5 rather than a distinct model.

Widen it slightly and it gets worse. ARC Prize runs three benchmarks on one page and the tab you land on is not the one people quote: ARC-AGI-3 leads with Claude Opus 5 (High) at 30.16%, ARC-AGI-2's top AI system is GPT-5.6 Sol (Max) at 93% behind a Human Panel row at 100%, and on ARC-AGI-1 three AI systems tie at 98%. The Humanity's Last Exam page ships two rankings at once: the live board at the top says Fable 5, 52.72%, and a static table further down the same page says Gemini 3 Pro, 38.3%, under a caption reading "Judge Model: o3-mini | Dataset Updated: April 3rd, 2025." Counting every sub-board on those six sites gives nine rank-1 rows, nine distinct names, five distinct base models.

Which boards publish an error bar, and which do not?

Three of the seven publish one, three publish none at all, one I could not read. On all three that publish, the rank-1 to rank-2 gap sits inside the interval:

• Arena text. Rank 1 is 1506.49, rank 2 is 1504.64: a gap of 1.84 Elo against a rank-1 interval of plus or minus 5.41. Arena publishes a rank spread column that does this work for you: four models carry an upper rank bound of 1. Across 391 models and 7,779,985 votes, the top five span 9.26 Elo.
• GDPval-AA v2. Rank 1 is 1849, stated -22 / +22; rank 2 is 1817, stated -21 / +21. A 32-point gap, and 1827 still overlaps 1838.
• SWE-bench Pro public. Rank 1 is 61.50 ±3.10, rank 2 is 59.10 ±3.56. A 2.40-point gap inside intervals wider than the gap.

Two things I will not claim. Overlapping intervals are not a significance test: non-overlap would imply a difference, overlap does not prove there is none. And these intervals are not comparable across boards. Arena Elo and GDPval-AA Elo share a name and nothing else, so ±5 on one and ±22 on the other say nothing about which board is noisier.

For the three boards with no error term, nobody can say whether their gaps are meaningful. Artificial Analysis prints an integer index with no interval, and its rank 1 and rank 2 both score 63 out of 159 models ranked. That is not a statistical tie. It is a rounding tie, a smaller claim.

Which board could I not read?

OSWorld-Verified. It renders client-side, and a raw fetch returns the literal strings "Loading verified benchmark data..." and "Updated: --" where the numbers should be. The Google Sheet linked from that page is a task-annotation worksheet, not the leaderboard. I did not see a rendered value, so I am not reporting one.

ARC Prize renders client-side too, with zero table rows in its raw HTML. I got its numbers from the JSON the page itself loads, at arcprize.org/media/data/leaderboard/v1.json, v2.json and v3.json. Those files carry a generatedAt stamp of 2026-08-13T23:42:30Z, two days before my fetch. That is the real freshness of the ARC numbers above, visible only by opening the pipe instead of the page.

Why is "the boards disagree" the wrong way to say this?

Because it implies there is one quantity called "best" that they are each estimating badly. There is not.

Arena text measures which of two blind responses a human voter preferred. That is preference, not capability, and a model that is warmer or better formatted can win votes without being more correct. SWE-bench Pro measures whether a patch makes failing tests pass across 1,865 tasks in 41 repositories. GDPval-AA v2 grades work products on 220 tasks from 44 occupations. ARC-AGI targets skill acquisition on puzzles built to be easy for people. Humanity's Last Exam is closed-ended expert exam questions, the exam-taking half of intelligence that "PhD-level reasoning" is indexed to. Five different questions. Five different answers is correct behavior; agreement would mean four are redundant.

There are wrinkles underneath the scores. Arena flags three of its 391 rows "Preliminary." The top five rows on SWE-bench Pro's public board carry an asterisk resolving to "Run with mini-swe-agent harness," so the leaders ran a different scaffold than the models below. ARC's ARC-AGI-3 leader cost $20,657.37 to evaluate. None of that is fraud. All of it means the score column scores a system, a harness and a budget, not a model.

So how should you actually pick a model?

Stop reading rank. Rank compresses things you do not care about into one integer.

Do this instead. First, read only the eval whose task shape matches yours. If you ship code, SWE-bench Pro and its harness footnote beat any composite index. If you produce documents and slides, GDPval-AA v2 was built from occupational tasks and is closer to your job than an Elo score from anonymous voters. If your product is a chat surface, Arena's preference signal is the right instrument, and its rank spread column tells you honestly that four models are tied for first.

Second, check the column nobody screenshots. Every board above carries cost, and the spread is enormous: the ARC-AGI-3 leader ran to five figures while ARC-AGI-1 leaders sit near a dollar per task. A model you cannot afford at your volume is not your number one.

Third, run your own eval. Twenty prompts from your real backlog, graded by you, beats every public board for your decision, and it is the only measurement a leaked test set cannot contaminate. Public numbers have a documented tendency to be run by the vendor, and "which coding tool wins" is a comparison that mostly cannot be priced honestly either.

What does this count not prove?

It does not prove the boards are wrong or gamed. Every one publishes its method, and three publish uncertainty when most products publish none. It does not prove the models are equivalent: on ARC-AGI-3 rank 1 scores 30.16% and rank 2 scores 7.78%, a real and enormous gap. It shows only that at the top, on one dated day, the leaders are close enough and the measurements different enough that "number one" is a marketing position, not a fact.

And it is one day. To re-run it: open the seven URLs in the table, record the top row and its interval, count the distinct names. If you get a number other than six, that is the finding, and it is the same finding as mine.