APEX-Agents tests long-horizon agent execution on professional-services work: investment banking, consulting, and legal tasks, 480 of them across 33 simulated worlds. The tensor.news record carries 58 scored models on it, with Gemini 4 Carbon at the top on 84.2.
The benchmark grades A, 85 out of 100, 32nd of 63 tracked benchmarks. The score is earned on discrimination (100, on a 79.8-point spread), saturation (97.1), harness consistency (100) and age (100). One axis carries the whole result: contamination, at 3.4 out of 100 — the lowest of any A-grade benchmark on the site.
Every row on the APEX-Agents leaderboard carries a source of unknown. Not one names an evaluator. None is marked independently reproduced, and none is marked self-reported. All 58 sit in the unverified tier.
This is not a criticism of the benchmark, which is well documented and recent. It is a description of what the record can prove. APEX-Agents is one of 11 of the 63 tracked benchmarks where no score names an evaluator; the other ten are ARC-AGI-2, SimpleQA Verified, GDPval, SpireClimb, Furniture Assembly, CL-bench, CL-bench Life, PostTrainBench, ExploitBench, and Remote Labor Index. A reader asking "who ran this?" gets no answer on APEX-Agents, and gets no answer on those either.
The benchmark's own harness note in the record explains where the numbers come from: the Epoch export reports model, model version, performance, an optimised flag, a score source, the benchmark release date and the model date. On APEX-Agents, the score source field is empty on all 58 rows, and so is the optimised flag.
APEX-Agents was released in 2026-01-21. The record's contamination window ratio is 0.966: 56 of the 58 dated rows belong to models released after the benchmark went public. The remaining two — DeepSeek-V3.2 and gpt-oss-120b — predate it, so their evaluations necessarily ran later.
The record's own contamination history for this benchmark states the authors' position plainly: Released 2026-01; publicly available task prompts and files may overlap with web-scraped training data — flagged by authors.
That is a self-reported risk of train-set overlap, and it is why the contamination axis scores 3.6 while the privacy field reads unknown. The test set's privacy status is not established, and the benchmark sits at the very top of the release-date exposure distribution. Every other benchmark with a public test set scores 0 on this axis because the exposure ratio is 1.0 — APEX-Agents scores 3.4 because 96.6% of its rows postdate the benchmark.
This is the one place where a benchmark's own documentation and the record's evidence tiers point the same way, and the A grade does not reflect it. The grade is a weighted sum, and the four other axes carry 85% of the weight.
APEX-Agents has 58 rows and 17 priced ones. The first priced row is at rank 16. None of the top 10 carries a price, and only 2 of the top 20 do. In the bottom half of the leaderboard, 12 of 29 rows are priced.
So the shape of the board and the shape of the price table are opposite. A buyer who opens the leaderboard to find the strongest agent and then tries to price it finds an empty cell at exactly the point where the decision gets expensive.
The price-per-point ranking reverses the leaderboard just as sharply. The most efficient priced model is DeepSeek-V3.2 at 21.3 for $0.40 per million output tokens — 53.25 points per $1M — sitting at rank 57 of 58, the second-worst model on the board. The top model, Gemini 4 Argon, is unpriced. Among the priced rows, Claude Fable 5 costs $50 per million output tokens for 63.6 points (1.27 points per $1M, corroborated), while DeepSeek-V3.2 delivers 53.25 for $0.40.
The record's metric field for APEX-Agents reads: normalized performance in Epoch CSV, rendered as accuracy (%). That is a two-step transform, and only the second step is visible on the page.
The underlying export carries raw performance values as fractions between 0 and 1 — the leaderboard's top row, Gemini 4 Carbon, is 0.842 in the export and 84.2 on the site. The normalization step from a fraction to the published score is not documented on the record, so the reader cannot tell how many of the 480 tasks a point corresponds to. The distribution itself shows how coarse the scale is: of 58 rows, 12 fall between 20 and 39, 27 between 40 and 59, 15 between 60 and 79, and three rows sit above 80. The record does not state the task count behind a single point.
The benchmark's own limitation note adds the shape of the ceiling: frontier Pass@1 scores were below 25% as of early 2026 against roughly 40% for Pass@8, closed-source models dominate the board, and world-environment fidelity limits how far the result travels outside the simulation. The tasks are set in simulated worlds authored by professionals with 5 to 10 years of experience at firms like McKinsey, Goldman Sachs, and Latham & Watkins, and are estimated at about 1.8 hours for a human professional each. No aggregate published human Pass@1 exists, so the record's human-baseline field carries the estimate and the absence together.
The score itself is normalized performance from the Epoch export, rendered on the site as accuracy (%). It is not a raw pass rate, and the record does not state how many of the 480 tasks a given point corresponds to. The benchmark's own limitation note is worth reading alongside it: frontier Pass@1 scores were below 25% as of early 2026 against roughly 40% for Pass@8, closed-source models dominate the board, and world-environment fidelity limits how far the result travels outside the simulation.
There is no aggregate published human Pass@1 either. The benchmark states that tasks are estimated at about 1.8 hours for a human professional, and that no human score is published — so the record's human-baseline field reads accordingly.
What the number supports: a model's normalized score on 480 simulated professional-services tasks. What it does not support: naming who ran any of the 58 evaluations, claiming the task files were unseen by the models trained after January 2026, or reading the leaderboard's top as a priced recommendation. It is a capability ordering with an open provenance question and a self-reported contamination risk, and the record says both things where a reader can see them.