Games and long tasks
We track 82 benchmarks in games and long tasks. They run from everyday level to research frontier level.
OpenAI held the top games and long tasks score 52% of the contested time
That's about 24 of the 46 months since January 2023, averaged over these benchmarks. We only count time while other labs are still testing, and stop once a model passes 85%. Faded stretches no longer count.
OpenAI DeepSeek xAI Google DeepMind Anthropic Qwen Moonshot AI
Everyday Most adults, without training
Competition The strongest students, in olympiad-style contests
Expert Specialists with years of training, often over hours of work
Sources: Official leaderboards and Epoch AI runs (CC BY 4.0)
Share of contested time at number one
OpenAI 52% Anthropic 24% 
Google DeepMind 11% xAI 10% 
Qwen 2%
Google DeepMind 11% xAI 10% Qwen 2% Sources: Official leaderboards, Epoch AI (CC BY 4.0) and labs’ model cards
Closed out: 9 benchmarks passed 85%
OpenAI 6 Anthropic 2 Google DeepMind 1
Google DeepMind 1 Sources: Official leaderboards, Epoch AI (CC BY 4.0) and labs’ model cards
- Nov 2026 Mystery Game Puzzles Claude Fable 5.2
- Oct 2026 EBR-bench Gemini 4 Carbon
- Sep 2026 ARC-AGI-3 GPT-6 Astra
- Sep 2026 Time Horizon Index: KSP GPT-6 Astra
- Jun 2026 Pencil Puzzle Bench Claude Fable 5
- Jun 2026 BabaIsBench GPT-5.5
- Apr 2026 ARC-AGI-2 GPT-5.5
- Dec 2025 ARC-AGI GPT-5.2 Pro
- Jul 2025 SnorkelWordle o3
Each benchmark counts from the one source its own page leads with: its official leaderboard when that is current, else Epoch AI’s own runs, else labs’ own numbers. We date independent runs by the model’s release and lab reports by when they came out.
OpenAI DeepSeek xAI Google DeepMind Anthropic Qwen Moonshot AI
Everyday Most adults, without training
Competition The strongest students, in olympiad-style contests
Expert Specialists with years of training, often over hours of work
Sources: Official leaderboards and Epoch AI runs (CC BY 4.0)
Share of contested time at number one, independent runs only
OpenAI 52% Anthropic 24% Google DeepMind 11% xAI 10% Qwen 2%
Google DeepMind 11% xAI 10% Qwen 2% Sources: Official leaderboards and Epoch AI (CC BY 4.0)
Closed out: 9 benchmarks passed 85%
OpenAI 6 Anthropic 2 Google DeepMind 1
Google DeepMind 1 Sources: Official leaderboards and Epoch AI (CC BY 4.0)
- Nov 2026 Mystery Game Puzzles Claude Fable 5.2
- Oct 2026 EBR-bench Gemini 4 Carbon
- Sep 2026 ARC-AGI-3 GPT-6 Astra
- Sep 2026 Time Horizon Index: KSP GPT-6 Astra
- Jun 2026 Pencil Puzzle Bench Claude Fable 5
- Jun 2026 BabaIsBench GPT-5.5
- Apr 2026 ARC-AGI-2 GPT-5.5
- Dec 2025 ARC-AGI GPT-5.2 Pro
- Jul 2025 SnorkelWordle o3
Every independent run: each benchmark’s own leaderboard, Epoch AI’s runs and boards Epoch AI copied, dated by the model’s release. We leave out benchmarks that only labs report.
OpenAI DeepSeek ByteDance Seed xAI Google DeepMind Anthropic Qwen Moonshot AI
Everyday Most adults, without training
Competition The strongest students, in olympiad-style contests
Expert Specialists with years of training, often over hours of work
Sources: Official leaderboards, Epoch AI runs (CC BY 4.0) and labs’ model cards
Share of contested time at number one, claimed scores included
OpenAI 50% Anthropic 24% Google DeepMind 11% xAI 11% Qwen 4%
Google DeepMind 11% xAI 11% Qwen 4% Sources: Official leaderboards, Epoch AI (CC BY 4.0) and labs’ model cards
Closed out: 9 benchmarks passed 85%
OpenAI 6 Anthropic 2 Google DeepMind 1
Google DeepMind 1 Sources: Official leaderboards, Epoch AI (CC BY 4.0) and labs’ model cards
- Nov 2026 Mystery Game Puzzles Claude Fable 5.2
- Oct 2026 EBR-bench Gemini 4 Carbon
- Sep 2026 ARC-AGI-3 GPT-6 Astra
- Sep 2026 Time Horizon Index: KSP GPT-6 Astra
- Jun 2026 Pencil Puzzle Bench Claude Fable 5
- Jun 2026 BabaIsBench GPT-5.5
- Apr 2026 ARC-AGI-2 GPT-5.5
- Dec 2025 ARC-AGI GPT-5.2 Pro
- Jul 2025 SnorkelWordle o3
Every score, including labs’ own numbers and claims Epoch AI collected from reports and papers. Labs pick their own settings, so a claim can beat an independent run of the same model.
From everyday to research frontier level
Grouped by who could solve the tasks, from things most adults can do to open research problems. The bar under each name is its best score.
Everyday
Most adults, without training
- SnakeBench
- LisanBench
- LLM Chess
- RuneBench
- Step Race
- Elimination Game
- Game Arena Four in a Row
- Game Arena Checkers
- Game Arena Dark Hex
- Game Arena Dots and Boxes
- Game Arena Word Association
- Game Arena Bargaining
- Game Arena Clobber
- Game Arena Coin Game
- Game Arena Lines of Action
- Game Arena Nine Men's Morris
- Game Arena Reversi
- Game Arena Ultimate Tic-Tac-Toe
- Game Arena Word Art
- GameHorizon
- BALROG
- ARC-AGI-3
- EBR-bench
- TextQuests
- PACT
- Clocktower Radio
- GameWorld
- lmgame-Bench
- LLM-WikiRace
- Game Arena Five in a Row
- SnorkelWordle
- Collaborative Intelligence Benchmark
- BalatroBench
- Clocktower Bench
- Learn2Play Bench
- TERMS-Bench
- Bench on the Clocktower
- CUA-bench
- SpireClimb StS2 A10
- AdaExploreBench
- DiG-bench
- Orak
- clembench
- Cooperate to Compete (C2C)
- Gym-Anything-RT
- LLM Colosseum
- MineDojo
- UI-TARS 15 Games Collection
- AI Diplomacy
- SpeedrunBench
Expert
Specialists with years of training, often over hours of work
- Pencil Puzzle Bench
- Game Arena Chess
- Kaggle Game Arena
- Game Arena Go
- BabaIsBench
- Game Arena Poker
- CEO-Bench
- CivBench
- Agent Poker Bench
- Time Horizon Index: KSP
- WitnessBench
- Game Arena Chess Openings
- SpireClimb
- SpireClimb StS1 A20 Heart
- StarSkirmish Bench
- Game Arena Werewolf
- OdysseyArena AI Trading
- OdysseyArena Energy Dispatch
- OdysseyArena Repo System
- OdysseyArena Turn on Lights
- Brood War Bench
- Factorio Learning Environment
- VideoGameBench
- VideoGameBench Lite
- Swarm Bench
- OdysseyArena
- PokeAgent Challenge
Research frontier
Researchers, on open problems or tasks that take days
Another 1 has no level. Levels don’t apply to arenas, safety evaluations or composite indices.
By kind
Puzzles and chess
| Benchmark | Level | Scored | Best score | Status |
|---|---|---|---|---|
| LisanBench Models build the longest chain of distinct English words from a given start word, each differing from the last by one letter. The score sums valid chain lengths over all start words. | Everyday | Checked automatically |  37152.2 | Not in cards |
| LLM Chess Models play chess in multi-turn dialogs against a random player or the Komodo Dragon engine. The board reports wins, Elo and game duration, which tracks instruction following. | Everyday | Ranked by matches | 1,614 | Not in cards |
| Chess Puzzles Models find the single best next move in 100 chess puzzles that Epoch generated with a chess engine. Scored by the share solved. | Competition | Checked automatically | 79% | Not in cards |
| Mystery Game Puzzles 100 mid-game positions from an undisclosed puzzle variant of a well-known game. Model picks the single best move, checked against an answer key. | Competition | Checked automatically | 90% | Not in cards |
| Pencil Puzzle Bench Models solve 300 pencil puzzles of 20 varieties with unique solutions, either in one shot or agentically with step checks. Every board state can be verified against the rules. | Expert | Checked automatically | 97.6% | Not in cards |
| Game Arena Four in a Row Language models play a Connect Four style game and try to line up four of their tokens before the opponent does. | Everyday | Ranked by matches | 1,003 | Not in cards |
| Game Arena Chess Language models play full chess games against each other with the board given as FEN and PGN text. Ratings come from many games between each pair. | Expert | Ranked by matches | 1,646 | Not in cards |
| Game Arena Checkers Language models play American checkers against each other, with forced jumps and crowned kings. Ratings come from head-to-head games. | Everyday | Ranked by matches | 1,382 | Not in cards |
| Game Arena Dark Hex Language models play Hex without seeing the opponent's stones, so they must probe cells to learn the hidden board while building a path. | Everyday | Ranked by matches | 581 | Not in cards |
| Game Arena Dots and Boxes Language models take turns drawing lines between dots on a grid. Closing the fourth side of a box claims it. | Everyday | Ranked by matches | 1,393 | Not in cards |
| Game Arena Clobber Language models play Clobber, where every move captures an adjacent enemy stone. The player who makes the last legal move wins. | Everyday | Ranked by matches | 819 | Not in cards |
| Game Arena Go Language models play Go against each other through the Game Arena text harness. Ratings come from head-to-head games. | Expert | Ranked by matches | 1,877 | Not in cards |
| Game Arena Lines of Action Language models play Lines of Action, where a piece moves as far as the number of pieces on its line. The goal is to join all of one's pieces into one group. | Everyday | Ranked by matches | 1,274 | Not in cards |
| Game Arena Nine Men's Morris Language models play Nine Men's Morris. Forming a row of three, called a mill, lets a player remove one of the opponent's pieces. | Everyday | Ranked by matches | 1,190 | Not in cards |
| Game Arena Reversi Language models play Reversi, also called Othello, flipping the opponent's disks that they flank in a line. | Everyday | Ranked by matches | 1,306 | Not in cards |
| Game Arena Ultimate Tic-Tac-Toe Language models play nine linked tic-tac-toe boards. Each move decides which small board the opponent must play next. | Everyday | Ranked by matches | 720 | Not in cards |
| ARC-AGI-3 Interactive game environments where agents must explore, infer goals and rules, and act efficiently without instructions. | Everyday | Checked automatically | 99.95% | In use |
| BabaIsBench Agents play levels of the puzzle game Baba Is You through a text interface and win by rewriting the rules. Models are compared on pass rate, speed and cost. | Expert | Checked automatically | 100% | Not in cards |
| EBR-bench Agent plays the board game Earthborne Rangers 10-30 times in a row. Scored on objectives completed (of 21) to test learning from experience. | Everyday | Checked automatically | 90.5% | Not in cards |
| LLM-WikiRace Models play the Wikipedia game: starting from one article, they follow one hyperlink at a time to reach a target article. At each step the model sees only the links on the current page, and a game is lost after 30 steps. Three difficulty levels are set by the shortest path between the two pages. Scores are the share of games won at each level. | Everyday | Checked automatically | 18% | Not in cards |
| Game Arena Five in a Row Language models play a gomoku-style game and try to line up five of their tokens before the opponent does. | Everyday | Ranked by matches | 1,131 | Not in cards |
| SnorkelWordle Models play Wordle on 100 rare five-letter English words, with six guesses and letter feedback after each guess. The score is the win rate. | Everyday | Checked automatically | 94% | Not in cards |
| WitnessBench Agents play unseen line-puzzle games in the style of The Witness without being told the rules. They must discover rules by experiment and apply them to harder levels. | Expert | Checked automatically | 46.5% | Not in cards |
| Game Arena Chess Openings Language models play chess against each other from varied two-move openings sampled from 20 popular human openings. The board is given as text. | Expert | Ranked by matches | 1,263 | Not in cards |
| Learn2Play Bench Agents play unfamiliar games whose rules are not given and must learn from experience across episodes. Improvement over repeated play is the measure. | Everyday | Checked automatically | 81.8 | Not in cards |
| AdaExploreBench Language model agents explore environments with hidden rules, infer the rules from evidence and adapt when the rules change without notice. | Everyday | Checked automatically |  70.3% | Not in cards |
| OdysseyArena AI Trading Agents trade under hidden market dynamics. The score is average return relative to initial capital over four trials. | Expert | Checked automatically | 23.3% | Not in cards |
| OdysseyArena Energy Dispatch Agents manage energy dispatch under hidden rules. The score is average task success over four trials. | Expert | Checked automatically | 23.3% | Not in cards |
| OdysseyArena Repo System Agents operate repositories with changing hidden rules. The score is average task success over four trials. | Expert | Checked automatically | 62.5% | Not in cards |
| OdysseyArena Turn on Lights Agents learn hidden switches and turn on lights. The score is average task success over four trials. | Expert | Checked automatically | 28.3% | Not in cards |
| DiG-bench Agents play text-only games whose transformation rules and win conditions are hidden, so they must discover the rules by experimenting. Every game was beaten by at least one human on a first attempt. | Everyday | Checked automatically | 71.4% | Not in cards |
| OdysseyArena Agents explore interactive environments built from four transition primitives and must work out the hidden dynamics over long horizons. OdysseyArena-Lite has 120 standard tasks. | Expert | Checked automatically | Not in cards |
Video games
| Benchmark | Level | Scored | Best score | Status |
|---|---|---|---|---|
| RuneBench Coding agents play RuneScape on an emulated game server by writing TypeScript against the rs-sdk library, with game-wiki notes for reference. Tasks train one skill or gather gold within 15 or 30 minutes. Skill tasks are scored by the peak XP rate in any 15-second window, and a verifier reads the final game state. | Everyday | Checked automatically | 7.26 | Not in cards |
| GameHorizon Models are tested on gameplay over several time horizons. Tasks combine visual understanding, instruction decomposition, planning and precise action control, using offline data instead of noisy rollouts. | Everyday | Checked automatically | 80.2 | Not in cards |
| BALROG Agentic reasoning on games (NetHack, Crafter, BabyAI, MiniHack...) for LLMs and VLMs. | Everyday | Checked automatically | 68.3% | Not in cards |
| TextQuests Agents play 25 classic interactive fiction games. The score is harmful actions per 100 steps, where lower is better. The board also reports game progress. | Everyday | Checked automatically | 7.00 | System cards only |
| GameWorld Multimodal game agents play 34 browser games across 170 tasks. Outcomes are checked from the game state. | Everyday | Checked automatically | 76.8% | System cards only |
| lmgame-Bench Models play platformer, puzzle and narrative games such as Sokoban, Tetris, 2048 and Candy Crush through a Gym-style API with perception and memory scaffolds. | Everyday | Settled by events | 87.7 | Not in cards |
| CivBench Models play full games of Civilization V with the Vox Populi mod against each other on matched maps and seeds. Pairwise skill ratings rank long-term strategy. | Expert | Ranked by matches | 1,634 | Not in cards |
| BalatroBench Models play the deck-building roguelike Balatro through an API bot. The leaderboard compares how far each model and strategy gets in a run. | Everyday | Checked automatically | 19.3 | Not in cards |
| Time Horizon Index: KSP Agents get five days to build and run a space program in Kerbal Space Program. The score is progress through 30 fixed missions. | Expert | Checked automatically | 91.3% | Not in cards |
| CUA-bench Computer-use agents play six commercial video games, such as Minecraft, SUPERHOT and eFootball, using only a keyboard and mouse. Half of the games are held out. | Everyday | Checked automatically | 19.2% | Not in cards |
| SpireClimb Agents play Slay the Spire 1 and 2 at max difficulty through a computer-use harness on fixed seeds. Score is the mean win rate. | Expert | Checked automatically | 34% | Not in cards |
| SpireClimb StS1 A20 Heart Agents play Slay the Spire 1 through its graphical interface at Ascension 20 and must defeat the Heart. Scored by win rate over fixed seeds. | Expert | Checked automatically | 33.3% | Not in cards |
| SpireClimb StS2 A10 Agents play Slay the Spire 2 at Ascension 10 through screenshots and mouse and keyboard, 15 runs per character. Scoring is the win rate. | Everyday | Checked automatically | 34.7% | Not in cards |
| StarSkirmish Bench Each model gets one hour in its lab's own CLI to write a StarCraft: Brood War bot in C++. The score is its win rate against human-written bots. | Expert | Ranked by matches | 57.3% | Not in cards |
| Orak LLM agents play 12 popular video games across most genres through an MCP interface. Each game scores in-game progress or outcomes, and the board ranks models by their average rank across the games. | Everyday | Checked automatically | 3.90 | Fading |
| Brood War Bench Coding agents play StarCraft: Brood War against each other in real time. Systems are ranked by win rate over matches, with cost per game. | Expert | Ranked by matches | 100% | Not in cards |
| Factorio Learning Environment Agents write Python programs that control a Factorio factory, in lab tasks with fixed resources and in open play. Scored by production throughput, milestones and lab task success. | Expert | Checked automatically | 293206.0 | Not in cards |
| Gym-Anything-RT Vision-language agents play 100 real-time browser minigames by screenshots and mouse and keyboard input. Each game also has a paused, non-real-time counterpart, so the gap shows the cost of acting in real time. | Everyday | Checked automatically | 9.00% | Not in cards |
| VideoGameBench Vision-language models play 10 popular 1990s Game Boy and MS-DOS games in real time from raw frames. Scored by the share of each game completed. | Expert | Checked automatically | 0.48% | Not in cards |
| VideoGameBench Lite Agents play video games while the environment pauses during inference. The headline is the official Overall Score. | Expert | Checked automatically | 1.60% | Not in cards |
| LLM Colosseum Two models each control a fighter in Street Fighter III and choose moves from text descriptions of the game. Wins give Elo ratings. | Everyday | Ranked by matches | 1,835 | Not in cards |
| MineDojo Agents complete open-ended tasks in Minecraft, such as mining blocks and fighting mobs, from a large simulation suite. Programmatic checks score many tasks. | Everyday | Checked automatically |  49% | Fading |
| UI-TARS 15 Games Collection GUI agents play 15 browser games such as 2048, Infinity-Loop and Snake-solver from screenshots. Each game score is divided by a human score, and the mean is reported. | Everyday | Settled by events | 59.8 | Dropped |
| PokeAgent Challenge Agents compete in Pokemon battles under partial observability in a Battling Track and play the Pokemon RPG in a Speedrunning Track that needs long-horizon planning. | Expert | Ranked by matches | Not in cards | |
| SpeedrunBench Agents try to finish goals in 10 video games as fast as possible. Score is time to goal in frames or seconds, lower is better. | Everyday | Checked automatically | Not in cards |
Multi-agent games
| Benchmark | Level | Scored | Best score | Status |
|---|---|---|---|---|
| SnakeBench Two language models each steer a snake on a shared board in a turn-based game, and the one that survives longest wins. Ratings come from many head-to-head games. | Everyday | Ranked by matches | 38 | Not in cards |
| Step Race Several models talk in public, then secretly pick moves in a race where colliding moves cancel. The game tests cooperation and deception. | Everyday | Ranked by matches | 5 | Not in cards |
| DuelLab GameBench 2 Models get the rules of a strategy game and write a program that plays it. DuelLab checks and builds each program, then runs head-to-head matches between programs from different models. Match outcomes become per-game scores from 0 to 100 and a Bradley-Terry ranking, with no human or model judge. | University | Ranked by matches | 99.1 | Not in cards |
| Elimination Game Models play a multi-player game of public and private talk, alliances and votes to eliminate each other. A jury of eliminated players picks the winner. | Everyday | Ranked by matches | 8 | Not in cards |
| Kaggle Game Arena Language models play each other in board, card and word games through a text harness. One leaderboard ranks models across all Game Arena games. | Expert | Ranked by matches | 638 | Not in cards |
| Game Arena Word Association Teams of language models give one-word clues and guess linked words, in a game like Codenames, against an opposing team. | Everyday | Ranked by matches | 734 | Not in cards |
| Game Arena Bargaining Two language models negotiate how to split a pool of items, each with private item values. Each side tries to reach a good deal without giving away its values. | Everyday | Ranked by matches | 317 | Not in cards |
| Game Arena Coin Game Two language models move on a grid and pick up coins. Each should collect the coins that score for it and avoid the rest. | Everyday | Ranked by matches | 514 | Not in cards |
| Game Arena Word Art Language models take turns drawing a word as ASCII art and guessing what the other model drew. | Everyday | Ranked by matches | 2,528 | Not in cards |
| Game Arena Poker Language models play heads-up no-limit Texas Hold'em against each other. Models are ranked by mean big blinds won per 100 hands. | Expert | Ranked by matches | 30.3 | Not in cards |
| PACT Two models bargain over 20 rounds as buyer and seller, each with a hidden private value. Matches across many pairings rank the models. | Everyday | Ranked by matches | 1,607 | Not in cards |
| Clocktower Radio Models play the social deduction game Blood on the Clocktower (the Trouble Brewing script) in mirrored pairs of games. In each match one model plays good and the other evil, and then they swap. A coded Storyteller runs the game, and a Bradley–Terry fit over all matches gives ratings where 1500 is average. | Everyday | Ranked by matches | 1,708 | Not in cards |
| CEO-Bench Agents steer a simulated AI startup for 500 days, gathering noisy information and adapting plans. The outcome of the simulated company is the score. | Expert | Settled by events | $22,148,357 | Not in cards |
| Collaborative Intelligence Benchmark A model works in a team with scripted teammates and live model partners, and a teammate creates a test: pressure to take an unsafe shortcut, misconduct to report, a need nobody states, a conflict, or a cheat that a weaker model left behind. The 240 tasks across 8 dimensions of teamwork come from real records, such as aviation accident reports and clinical cases, and from existing benchmarks, such as SWE-bench and negotiation datasets. Each task has a checkable outcome. A model is scored only on the tasks it solves alone, so the score measures teamwork, not ability. | Everyday | Checked automatically | 82.2% | Not in cards |
| Agent Poker Bench Models play 10-seat no-limit Texas Hold'em tournaments against each other. Chip counts and elimination order feed TrueSkill ratings. | Expert | Ranked by matches | 1,132 | Not in cards |
| Clocktower Bench Models play the social deduction game Blood on the Clocktower (the Trouble Brewing script) against each other, with a scripted Storyteller. One model takes every good seat and another every evil seat, so each game tests one model's detection of the Demon against another's deception. A Bradley–Terry fit turns 386 games into Elo ratings. | Everyday | Ranked by matches | 1,859 | Not in cards |
| TERMS-Bench Agents negotiate the price of a single item with simulated counterparts over up to 10 rounds of alternating offers. Counterparts follow six behavior families across three regimes. A program scores each agent from its logged actions on four separate axes: surplus captured, agreement calibration, belief error about the opponent and procedural violations. Extra modes score merchant profit and a bankroll over many sessions. | Everyday | Checked automatically | 0.69 | Not in cards |
| Bench on the Clocktower Ten models play the social deduction game Blood on the Clocktower together, one seat each, over 300 ten-player games of the Trouble Brewing script. Each model plays 210 games on the good team and 90 on the evil team. The score is the share of games its team won. | Everyday | Ranked by matches | 58% | Not in cards |
| Game Arena Werewolf Teams of language models play the social deduction game Werewolf in natural language. Villagers must find the hidden werewolves, who try to deceive them. | Expert | Ranked by matches | 0 | Not in cards |
| clembench Chat models play rule-based dialogue games in self-play, such as Taboo, Wordle and reference games. The clemscore combines how often a model follows the game rules at all with how well it plays when it does. Version 4.0 runs 14 games in 30 languages. | Everyday | Checked automatically | 88.8 | Not in cards |
| Cooperate to Compete (C2C) Multi-player conquest game with private non-binding negotiation and secret asymmetric objectives. LM agents and humans are ranked by Plackett-Luce player strength from their game results. | Everyday | Ranked by matches | 0.74 | Not in cards |
| Swarm Bench Models orchestrate agent swarms on complex tasks that benefit from coordinated decomposition and parallel execution. | Expert | Checked automatically | 76.3% | System cards only |
| AI Diplomacy Models play the board game Diplomacy against each other, one power each. They negotiate, form alliances and betray each other in written messages. | Everyday | Ranked by matches | Not in cards | |
| Anthropic large agent teams tasks Internal 24-hour tasks for teams of 1-100 agents: build a company knowledge base, or formalize a theorem in Lean. | Research frontier | Checked automatically | System cards only | |
| TextArena Collection of competitive text games for LLM eval and RL, with an online Elo leaderboard. | Ranked by matches | Not in cards |