Iceberg squares
Area equals task count. Above the dashed line is what anyone can see. Below is what Mercor sells.
Read it as: the public part is about a tenth of the whole.
What each chapter shows, how to read every chart, the terms the page uses, and the numbers worth remembering. Five minutes, then the full page reads fast.
Mercor publishes benchmarks to earn trust and attention, then licenses much larger datasets to AI labs for training.
Everything else on the page explains how it builds those benchmarks, checks them, launches them and where the process slips.
| Chapter | Question it answers | What the visual shows | Remember |
|---|---|---|---|
| TopHero | What is the business? | Two squares drawn to scale: public tasks above a dashed line, licensable tasks below. | 240 vs 2,620 |
| TopSeven findings | What matters most? | Seven headline numbers. The red one is a contradiction in Mercor's own data. | 9% · −38 pts |
| 01Footprint | What does Mercor sell? | Four business lines and its scale claims. Red numbers disagree with each other. | 5M+ vs 4M+ |
| 02Catalog | What is on the leaderboard site? | 43 tiles colored by type, then a dot table of what each type of page shows. | 8 own, 35 others' |
| 03Extended trick | How does Mercor show fresh data matters? | Slope chart: model scores on a public benchmark fall on Mercor's held-out version. | ~84% → ~46% |
| 04Methodology spec | How standard are the benchmark write-ups? | The 11-field card every benchmark gets, and a bar of how many runs each uses. | 11 fields, 31 cards |
| 05Commercial model | How does the money work? | Licensable volumes to scale, a six-step sales ladder, and the data-buying program. | 12,000 tasks |
| 06Partners | What does "in partnership with" mean? | Three cards, each with a bar of who built how much. | 3 kinds of deal |
| 0712 quality gates | What happens before a benchmark ships? | Twelve numbered checks. The three outlined have the strongest numbers. | 98.5% · 20% |
| 08Launch playbook | In what order do things go public? | A timeline of six launches with a symbol per event, then the usual sequence. | arXiv 1–2 days early |
| 09Weaknesses | Where does it slip? | Eight red-flagged problems, mostly the same fact differing across pages. | 100 vs 160 tasks |
| 10Unknowns | What can't be seen from outside? | A table of gaps and who could fill each one. | 6 open questions |
Area equals task count. Above the dashed line is what anyone can see. Below is what Mercor sells.
Read it as: the public part is about a tenth of the whole.
One tile per benchmark page. Black: partner. Teal: Mercor-built. Pale teal: Mercor-extended. White: open-source run by Mercor.
The sea of white tiles is the point: most of the catalog is other people's benchmarks.
Five dots per cell. Filled dots show the share of pages of that type that include the feature. The number beside them is the exact count.
Only Mercor's own pages fill the uncertainty, FAQ and "who writes" rows.
Left dot: a model's score on the public benchmark. Right dot: its score on Mercor's held-out version. Each line is one model.
Steep downward lines mean the public score overstates how the model does on new tasks.
Each segment is how many of the 31 benchmarks use that number of runs per task. Wider means more benchmarks.
Most run each task 3 times, then average.
Bar length is the number of tasks Mercor licenses, on one scale up to 12,000. The teal sliver is the public set, where stated.
Compare bar lengths directly. They are to scale.
Six steps from a free leaderboard to a full research partnership. Taller steps mean a bigger commitment from the buyer.
Every step is reachable from one page.
Black is the partner's share of the build, teal is Mercor's. The split is illustrative, not measured.
Same label on the website, very different deals underneath.
Numbered in the order work happens. Each card has what Mercor does, the number behind it, and an evidence label.
An outlined card has the strongest proof.
Hollow circle: repo created. Diamond: arXiv paper. Teal dot: blog. Black square: open data. One row per launch.
A long grey line means open data came weeks after the launch.
A red "!" or a red number marks a problem, usually two Mercor pages that disagree about the same fact.
These are the places to double-check before quoting Mercor.
Re-read on the original source (Mercor's site, paper or Hugging Face page) on 12 Nov 2026. Safe to quote.
Found by a research agent that cited a primary source, but not re-read. Check the source before quoting.
Reasoning from patterns or web-archive captures. Treat as a lead, not a fact.
Both. Its benchmark worlds are mostly fictional (22 of 33 APEX-Agents worlds are fully fictional). Separately, its data-buying program pays companies for real workplace data, which it de-identifies and licenses to labs.
No. The public APEX-Agents set is CC-BY-4.0 but marked "exclusively for model evaluation". Training, fine-tuning and scraping are forbidden. Training rights come with a license.
They show models scoring much lower on tasks they have never seen. That is the argument for buying fresh, private data, made with one chart.
Very. On 12 Nov 2026 they already listed Fable 5.2, GPT-6.1 Sol, Gemini 4 Carbon, Grok 4.6 and Muse Wonder 1.0.
No. Every step past the free leaderboard goes through a "Talk to our team" form.
Consistency. Release dates, task counts and expert counts differ between Mercor's own pages, because the same fact lives in several places.