Poker: machine Libratus against human Four HUNL pros
Carnegie Mellon’s Libratus beat four elite heads-up no-limit specialists by 147 big blinds per 100 hands over 120,000 hands in January 2017, with duplicate hands dealt to remove card luck. Overtaq records a confirmed, directly comparable crossover — the first decisive machine win in a large imperfect-information game.
The record in full
| Discipline | Poker HEADS-UP NO-LIMIT TEXAS HOLD’EM |
|---|---|
| What is measured | Long-run win rate against elite professionals, measured in big blinds per 100 hands |
| Human benchmark | Four HUNL pros — Jason Les, Dong Kim, Daniel McAulay, Jimmy Chou, Invited heads-up specialists UNITED STATES · 30 JANUARY 2017 · BRAINS VS ARTIFICIAL INTELLIGENCE, RIVERS CASINO, PITTSBURGH |
| Machine benchmark | Libratus — Libratus 30 JANUARY 2017 · BRAINS VS ARTIFICIAL INTELLIGENCE, RIVERS CASINO, PITTSBURGH |
| Attributed to | Carnegie Mellon University UNITED STATES |
| Difference | No single comparable number. The machine passed the best humans in a contest with an unambiguous result. |
| Crossover | 30 January 2017 — Libratus vs four professionals Libratus beat four elite heads-up specialists by 147 big blinds per 100 hands over 120,000 hands. |
| Verification status | VERIFIED Verified |
| Comparability status | DIRECT Directly comparable |
| Overtaq classification | CONFIRMED A machine passed the human benchmark in a contest or measurement Overtaq judges both well evidenced and legitimately comparable. |
| Data last checked | 23 AUGUST 2026 |
Conditions and comparability
Machine conditions
120,000 hands over 20 days, with duplicate hands dealt to control for card luck. Win rate of 147 big blinds per 100 hands, statistically significant.
Why this is directly comparable
Directly comparable and unusually well controlled: identical rules, a pre-agreed hand count large enough for statistical significance, and duplicate hands dealt to both sides to remove card luck. Two years later Pluribus extended the result to six-player games, which is the harder and more commercially relevant format.
Machine progression
3 recorded results. Overtaq keeps every published result rather than only the current best, so the frontier can be read as a history. This table is the full data behind the chart on the interactive record.
| Date | Result | System | Status | Note |
|---|---|---|---|---|
| 24 April 2015 | Claudico loses to pros | Claudico Carnegie Mellon University |
REPORTED DIRECT | The predecessor system lost a 80,000-hand heads-up match, though not by a statistically significant margin. |
| 30 January 2017 | Libratus wins by 147 bb/100 CROSSING | Libratus Carnegie Mellon University |
VERIFIED DIRECT | The crossover, over 120,000 hands with duplicate dealing. |
| 11 July 2019 | Pluribus beats pros six-handed | Pluribus Carnegie Mellon University and Facebook AI |
VERIFIED DIRECT | Extended the result to multiplayer poker against elite professionals, including bracelet winners. |
Evidence
2 sources. Every source is a direct link to the primary record where one exists; where Overtaq has not been able to confirm an exact permanent URL, the entry says so rather than linking to a search page.
- Science 359, 418–424 — Superhuman AI for heads-up no-limit poker: Libratus beats top professionals
17 DECEMBER 2017 · DOI:10.1126/science.aao1733
The 147 bb/100 win rate, the 120,000-hand format, duplicate dealing, and the professionals involved.
- Science 365, 885–890 — Superhuman AI for multiplayer poker
11 JULY 2019 · DOI:10.1126/science.aay2400
Pluribus beating elite professionals in six-player no-limit hold’em.
Social cards
Every record is rendered into four formats in both themes from the same data, so the verification and comparability status travels with the numbers wherever the card is posted.
