Fifteen models score a perfect 100 on OTIS Mock AIME 2024-2025. Only four of them have a published price. Those four range from $20 to $180 per million output tokens — a 9-times spread for exactly the same score.
This is the clearest illustration in the record of what a saturated benchmark does to a leaderboard. There is no ranking among the fifteen. There is a tie, and inside that tie a price axis that the leaderboard itself does not show.
OTIS Mock AIME 2024-2025 is a set of 45 competition-math problems: 15 from Mock AIME 2024, plus 15 each from Mock AIME 2025 I and 2025 II. Each problem has an integer answer between 0 and 999, and the exam runs three hours. The problems were authored for the benchmark by Evan Chen's OTIS program — Olympiad Training for Individual Study — rather than republished from official AIME papers, because the official papers have circulated for decades. The benchmark was released on 2024-12-19 and is licensed CC BY 4.0.
The design goal is frontier reasoning rather than recall. Evan Chen's own estimate is that the Mock AIMEs run 2 to 4 problems harder than the official AIME, and that an AIME qualifier threshold of 7-to-10 correct out of 15 — roughly 50% — is the reference point for a human. Problems near IMO difficulty appear in the set, with some artistic license; image-based problems are excluded so every problem stays text-parity comparable.
Epoch AI's Capabilities Index reports the scores, and this record reads from that CSV. Epoch is the source for 126 of the 169 rows; the other 43 carry no recorded source. The harness flag is therefore the strongest in the benchmark set — no mixed-source penalty, and a harness comparability of "consistent harness" — because the 126 sourced rows all come from one evaluator in one harness. The benchmark earns a Benchmark Integrity grade of B at 80 out of 100, 42nd of 63 tracked benchmarks.
Fifteen rows carry a score of 100. Ranked by the leaderboard's own ordering, they are: Gemini 4 Argon, Gemini 4 Carbon, Claude Fable 5, Claude Fable 5.1, Kimi K3.1 Pro, Qwen3.8 Max (0902), Claude Opus 5.5, Claude Sonnet 5.5, GPT-5.5, GPT-5.5 Pro, GPT-5.6 Sol, GPT-6 Astra, GPT-6 Sol, GPT-6.1 Sol, and GPT-6.1 Terra.
Eleven of those fifteen carry no source at all — their provenance is recorded as unknown — and the same eleven have no published price. So the set of models you can both price and verify is four: Claude Fable 5, GPT-5.5, GPT-5.5 Pro, and GPT-5.6 Sol. All four have independently reproduced scores from Epoch, and all four are marked as optimized runs, which is the test-time-compute setting most labs publish in their own model cards.
That leaves a leaderboard whose top fifteen is split cleanly: four rows with a price and a reproduction, eleven rows with neither. There is no row in the fifteen that has a price without a reproduced score, and no row with a reproduced score that lacks a price. The split is exact.
One detail sharpens it. Every one of the 126 Epoch-sourced rows carries an optimized flag — test-time compute on — and every one of the 43 rows with no source carries no such flag. The two sets are exactly coextensive, so the flag is a marker for "this row came from Epoch", not independent evidence about how a model was run. The priced four are Epoch rows; the unpriced eleven are not.
| Model | Score | Output $/M | Input $/M | Blended $/M | Points per $1M output | Price sources |
|---|---|---|---|---|---|---|
| GPT-5.6 Sol | 100 | $20 | $4 | $8.00 | 5.00 | single-source |
| GPT-5.5 | 100 | $30 | $5 | $11.25 | 3.33 | corroborated |
| Claude Fable 5 | 100 | $50 | $10 | $20.00 | 2.00 | corroborated |
| GPT-5.5 Pro | 100 | $180 | $30 | $67.50 | 0.56 | corroborated |
The output price range is $20 to $180, a 9.0-times spread. The blended range, weighting input at three times output, is $8.00 to $67.50, an 8.4-times spread. Every one of these models records the same score on the same 45-problem set.
If you buy tokens and this benchmark is your target, the tie is decided entirely by the invoice. GPT-5.6 Sol returns 5.00 points of OTIS Mock AIME per million output tokens; GPT-5.5 Pro returns 0.56. One caveat on the cheap end: GPT-5.6 Sol's price comes from a single source, while the other three are corroborated across sources, so the 9-times figure is a floor on the true spread rather than a symmetric claim.
The scores around the tie are compressed in a way that gives the tie away. Rank 16 sits at 99.72, rank 20 at 99.44, rank 22 at 99.17. The gap from 100 to 99.72 is 0.28 points; from 100 to 99.44 is 0.56 points. The entire distance from rank 1 to rank 20 is under one point.
That compression is a property of the metric, and it matters for how far you can read a difference. Epoch reports OTIS Mock AIME as a normalized performance, not a raw count of solved problems, and the source values are rational fractions — the top rows sit on denominators of 8,991 and 2,997, which is 999 multiplied by 9 and by 3. The finest gap anywhere in the 169-row distribution is 0.0111 points. So a 0.28-point gap is not a quarter of a problem; it is roughly 25 grid steps on a scale whose granularity the record does not otherwise explain, and the record does not say how many items that scale corresponds to.
What the record does say is that the scale's resolution is far finer than the differences it is being used to judge. The benchmark's own saturation flag makes the call: severity 0.201, a 34-model cluster bunched within 2 points of the top, and a ceiling proximity of 1.0. The spread score is 100 out of 100 because the full distribution runs from a perfect 100 down to a zero.
Here is the distribution that produces the tie: 15 models at 100, 32 more in the 95-to-99.99 band, 43 in the 80-to-94.99 band, 31 in the 50-to-79.99 band, 48 below 50, and one model that scored zero.
Contamination scores 21.9 out of 100 for OTIS Mock AIME, against 2.7 for GPQA diamond. The test set privacy is recorded as unknown, and the contamination window ratio is 0.781 — which means 78.1% of the dated rows come from models released after the benchmark's 2024-12-19 release date, so those models had the opportunity to train on the problems if they were public.
That is the mechanism the record uses, and it is worth being precise about: the ratio is a count of models postdating the release, not a measure of how long the test set has been available. OTIS scores better than GPQA diamond on this axis because 21.9% of its rows come from models that predate the release, while on GPQA diamond — released 2023-11-20 — almost every model in the record postdates it.
The benchmark's own design note is why this is a mild score rather than a severe one: the problems were novel and unreleased before 2024-2025, and Epoch's stated rationale for preferring authored sets over official AIME is exactly the leakage risk. The test set privacy stays unknown rather than public because nobody has established that the set stayed unpublished. The 21.9 out of 100 is a statement about how the record scores an unknown privacy state, not a finding that contamination happened.
The discrimination score is 100 out of 100 with a full 100-point spread, and the 126 Epoch-sourced rows share one harness, with no mixed-source flag and no mixed-optimization flag. The evaluation dates are current: 80 of the 169 rows were measured in 2026, 51 in 2025, 34 in 2024, and 4 in 2023, so this is a live leaderboard rather than a frozen archive.
The claim that 15 models tie at 100 is a real finding with a real limit. It means frontier competition-math ability has hit the ceiling of a 45-problem exam, and that on this benchmark price is the only remaining variable. It does not mean the 15 models are equally good at mathematics — the set runs from AIME-level to IMO-style difficulty, and a model that misses one hard problem may still be far better at everything below it.
Use OTIS Mock AIME as a ceiling detector, not a ranking. Once a benchmark produces a 15-way tie with 34 models clustered inside 3 points, the leaderboard has stopped ordering models and started reporting a threshold. The useful question it still answers is "which models clear a 45-problem competition-math exam", and on that question the answer is 15 models.
If you are choosing between the four priced models in the tie, the record gives you a clean 9-times price spread and nothing else to distinguish them: same score, same harness, same reproduction status. If you are choosing a benchmark to compare frontier models with, this one has run out of resolution — the next useful comparison has to come from a set larger than 45 problems, or from one whose items are harder than the models have trained on.
What this does not prove: that GPT-5.6 Sol is a better mathematician than GPT-5.5 Pro. It proves that on this benchmark they score identically, and that one costs nine times less per million output tokens.