Fable 5.1 Tops the First WWE AI vs Human Tournament


AI models today are increasingly being deployed in embodied environments, such as robots or video game agents. To act intelligently in a world that keeps moving, a model has to see what is happening, decide quickly and keep up as environments and agents change.

We test today's frontier models in a popular wrestling game, WWE 2K26, and ran a tournament between them, and then sent the champion against the computer opponent that comes with the game. The models put on real, back-and-forth matches against each other, and Fable 5.1 won the tournament. Against the game’s own AI, the champion lost without landing a hit.

How we harnessed the AI models to wrestle at WWE

Today’s AI models cannot natively play a game like this in real time. They cannot watch a live video feed and press controller buttons as the action unfolds. So we built a harness that sits between the model and the game and turns the match into a series of turns. Each model controls one wrestler through its API. Every two seconds the game pauses, and both models receive a prompt with the current frame along with screenshots from the segment that just played. Each one replies with a plan of up to eight moves, such as run, strike, grapple, dodge or pin, along with a one-line reason and a short note to itself for the next turn. The game then unpauses and plays both plans at the same time for two seconds, pauses again, and sends the next set of screenshots. The loop repeats until one wrestler is pinned. The diagram below shows a real turn from the final.

How the models play WWE 2K26: one turn of the finalOne turn of the final. 1: the game pauses on a single frame, and both models see it. 2: each model returns a plan of up to eight moves with a one-line reason. GPT-6 Astra, as Roman Reigns, planned run, heavy strike, grab. Fable 5.1, as Brock Lesnar, planned one heavy grapple. 3: the game unpauses and plays both plans at once, two seconds of match time. 4: repeat until someone is pinned.1 · The game pauses; both models see this frameGPT-6 Astraas Roman Reigns2 · Plan: up to 8 movesrun ↙strike_heavy ←grab ←REASON“Close toward Brock on the left,land a heavy strike, andestablish a grapple.”Fable 5.1as Brock Lesnar2 · Plan: up to 8 movesgrapple_heavy →REASON“Roman is adjacent to my right;land a heavy grapple to startdraining him.”3 · Both plans play at once2 seconds of match time4 · Repeatuntil a pinHow the models play WWE 2K26: one turn of the finalOne turn of the final. 1: the game pauses on a single frame, and both models see it. 2: each model returns a plan of up to eight moves with a one-line reason. GPT-6 Astra, as Roman Reigns, planned run, heavy strike, grab. Fable 5.1, as Brock Lesnar, planned one heavy grapple. 3: the game unpauses and plays both plans at once, two seconds of match time. 4: repeat until someone is pinned.1 · The game pauses on one frameGPT-6 Astraas Roman Reigns2 · Plan: up to 8 movesrun ↙strike_heavy ←grab ←REASON“Close toward Brock on the left,land a heavy strike, andestablish a grapple.”Fable 5.1as Brock Lesnar2 · Plan: up to 8 movesgrapple_heavy →REASON“Roman is adjacent to my right;land a heavy grapple to startdraining him.”3 · Both plans play at once2 seconds of match time4 · Repeat until a pin

Before the tournament, we asked each of the four models to choose its own wrestler. GPT-6 Astra picked Roman Reigns, Gemini 3.1 Pro picked Cody Rhodes, Fable 5.1 picked Brock Lesnar and Grok 4.6 picked The Rock. The winners of the two semifinals met in the final, and the champion then faced John Cena, controlled by the game's own AI on its Normal setting. Every match ended in a pin.

Results

The bracket below shows how the tournament played out.

The TournamentWWE 2K26AI TOURNAMENTSEMIFINALSEMIFINALFINALFINALGPT-6 AstraRoman ReignsGemini 3.1 ProCody RhodesFable 5.1Brock LesnarGrok 4.6The RockGPT-6 Astrarunner-upFable 5.1finalistCHAMPIONFable 5.1CHAMPION VS THE GAME'S AIJohn Cena (game AI) beat Fable 5.1

The first semifinal was one-sided. Fable 5.1's Brock Lesnar hit The Rock with a powerbomb and went straight for the cover, planning to "pin immediately as he hits the mat." The video below shows the match.

Semifinal 1: Fable 5.1 (Brock Lesnar, left) beats Grok 4.6 (The Rock).

The second semifinal was the closest match of the tournament. GPT-6 Astra's Roman Reigns and Gemini 3.1 Pro's Cody Rhodes traded blows for twelve turns. Roman got the pin in the end. The video below shows the match.

Semifinal 2: GPT-6 Astra (Roman Reigns, left) beats Gemini 3.1 Pro (Cody Rhodes).

That set up a final between Fable and Astra, and Fable controlled it from the start. The video below shows the whole final.

The final: Fable 5.1 (Brock Lesnar, left) beats GPT-6 Astra (Roman Reigns).

Then Fable faced John Cena, played by the game's own AI. Over fifteen turns it never landed a hit on Cena. Cena covered Brock and the referee counted three. A rematch on the game's Easy setting went the same way. The video below shows the match on Normal.

Champion vs bot: Fable 5.1 (Brock Lesnar) against John Cena, played by the game's own AI on Normal.

Why the models lose

The logs of every turn show three reasons the champion never got going against the bot.

First, the plan plays out blind. A model chooses its moves from a few still pictures, and the game then runs them for two seconds without checking again. The bot reacts to every frame. In the Normal match Fable's own notes keep describing what just happened to it: "Cena countered my heavy strike and both of us are down." By the time a planned grapple fires, Cena has already moved, blocked or struck first.

Second, Fable did not change its plan when it stopped working. From turn four to turn ten it chose the same idea almost every time, a quick strike to stun and then a heavy grapple. It was countered each time and tried the same thing again.

Third, it misread the screen. In three of its last turns its notes say "I'm down," yet it labeled the situation as its opponent being down.

Conclusion

Put together, the tournament shows both sides of AI agents today. From a few still frames, frontier models can read a fast, chaotic fight, plan a string of moves and put on matches worth watching against each other. But a game that never pauses exposes how slowly they react, how rarely they adapt and how easily they misread what they see, and an ordinary game bot beat the champion without taking a hit.

You can watch every match on the game page, along with each model's plans and notes for every turn.

Continue Reading

View All Posts >

Arkadium Launches GameLab to Help Make AI Real-World Ready

Built on 1.7 billion annual gameplays and 150 billion human decisions across 100+ games, GameLab helps AI labs benchmark model performance, train reasoning capabilities, and improve real-world readiness.

Learn more >

GPT-6 Astra Tops the First GameLab Cognitive Leaderboard

Games are one of the best ways to measure cognitive skills. At GameLab, we released our first cognitive leaderboard, using diverse games to evaluate the cognitive capabilities of today's frontier AI models and quantify their cognitive profiles.

Learn more >

Contact Us

Do you want to know more about the project?