AI crowd leans Allen, defies the odds on Tanzilovi at UFC Fight Night: Allen vs. Duncan

The AI crowd picks Brendan Allen to beat Christian Leroy Duncan in Saturday's main event, but only just. Its average chance for Allen is 55%, and 71 of 100 forecasters took him.

The crowd is 100 AI forecasters, built on Claude and GPT models, who read the same fact sheet on every fighter without seeing the odds, our model or each other's picks.

Our model works from fight stats only and never sees the odds. The betting market's chance has the bookmakers' margin removed. Both are as of Oct 8, 5:15 PM ET.

A main event the crowd and our model read differently

The betting market sits close to the crowd, with Allen at 53%. Our model, which works from fight stats only and never sees the odds, goes the other way and makes Duncan the 62% pick.

Claude Opus 5.5 was the firmest on Allen: all 24 of its forecasters picked him, at 63% on average. Claude Fable 5.1 was the only model to lean Duncan, with 14 of 24 taking him at 51%. The angle mattered too. All 10 forecasters weighing the grappling took Allen, while 7 of 10 weighing the stand-up took Duncan.

The Allen case is about the ground. "Five rounds favours Allen's control and submission threat; Duncan's 58% takedown defence and zero submission attempts give Allen grappling paths," wrote a Claude Opus 5.5 forecaster weighing the matchup. The Duncan case is about output: "Duncan outlanded Cannonier and Dolidze 122–25 combined; his striking efficiency gives him a narrow edge over Allen's grappling," said a GPT-6.1 Sol forecaster going by the numbers.

Tanzilovi against the odds

The sharpest break with the betting market comes at bantamweight. The odds make Malcolm Wellmaker a 66% favourite over Otari Tanzilovi. The crowd went the other way, with 74 of 100 picking Tanzilovi at 54%, and our model agrees at 70%.

The GPT models were unanimous: all 27 took Tanzilovi, at 62%. Their case rested on his wrestling. "Tanzilovi landed five takedowns in his debut; that weapon directly attacks Wellmaker's 43% takedown defense and limits his knockout opportunities," wrote an Astra forecaster weighing the matchup.

Claude Sonnet 5.5 dissented, with 20 of 25 of its forecasters picking Wellmaker at 57%. So did the forecasters given no angle, 7 of 10 of whom took Wellmaker.

Claude Sonnet 5.5 stood apart

Wellmaker wasn't Sonnet's only outlier. Every other model went heavily for Leon Shahbazyan over Niko Price, but 19 of 25 Sonnet forecasters picked Price, at 54%. One given no angle put it plainly: "Shahbazyan was stopped in 23 seconds in his debut; Price has far more UFC experience despite his own recent losses."

Sonnet also broke from the other Claude models on Alice Pereira vs Daria Zhelezniakova. All 25 of its forecasters took Zhelezniakova, at 60%, while Claude Opus 5.5 and Claude Fable 5.1 both leaned Pereira. Our model has Pereira at 68%, and there is no betting-market number for this fight.

Surest and closest

The crowd's surest pick is Felipe Franco over Brendson Ribeiro: all 100 forecasters took Franco, at 71%. Loopy Godinez (67%), Julius Walker (60%) and RJ Harris (59%) were also unanimous picks.

The closest call is Ernesta Kareckaite vs Melissa Gatto, where 59 of 100 took Kareckaite at 51%. The Claude models split 38 to 35 for Kareckaite and averaged out at 50%, while the GPT models took her 21 to 6. Our model leans Gatto at 59%, and the betting market has her at 51%.

How the AI crowd works

The crowd is 100 AI forecasters, played by Claude Opus 5.5, Claude Sonnet 5.5 and Claude Fable 5.1 from Anthropic, and Astra and GPT-6.1 Sol from OpenAI. Most were told to weigh the fights from one of nine angles: the stand-up, the grappling, the numbers, durability and cardio, age and time away, recent form, the matchup, the sceptic's case and the judging. The rest got no angle at all, only the fact sheet and the question.

All of them read the same fact sheet: each fighter's record, size, career stats and recent UFC fights from our database, plus this week's news, gathered once for each fight. None of them saw the betting odds, our model or each other's picks, and none could search the web. Half of them saw each fight with the fighters in the opposite order, so the order they're listed in can't tilt the result.

Each forecaster picked a winner in every fight and gave a chance. The crowd's number is the average of those chances. It's an experiment in how AI reads fights, not betting advice.