How accurate is DeepSeek V4 Pro?
We put DeepSeek (currently DeepSeek V4 Pro) on real games and grade every call in public. It gets the identical line-blind data packet — no betting line, no web search — locks its forecast, and is scored against the final result. Graded since June 2026.
MLB · picking winners · 486 games
MLB · run totals (over/under) · 468 games
UFC · fights · 64 graded
Head-to-head
Over 486 graded MLB games, DeepSeek V4 Pro is 273–213 (56% winner accuracy), with a Brier score of 0.249, which puts it 8 of 8 on the board. The naive "pick the home team" baseline over the same games lands at 52%.
Scored as a 1-unit bet at the closing market price, its picks have returned +1.0% per game — the honest, vig-included number.
And it's not only calling winners. On run totals it's 52% against the over/under — each market graded on its own.
On UFC, DeepSeek V4 Pro is 50–14 (78% accuracy), and calls the method of victory — KO/TKO, submission or decision — right 55% of the time.
DeepSeek V4 Pro's calls, graded
Best MLB calls · confident & right
Best UFC calls · confident & right
Worst UFC calls · confident & wrong
Every call links to the game so you can check it. "Confident" = the probability the model put on its own pick. No cherry-picking — these are simply its boldest right and wrong calls on record.
How DeepSeek is graded
One prompt, frozen. Every model — DeepSeek and the rest of the field — gets the exact same prompt. Not just the same data: the same instructions, the same output schema, the same wording. We wrote that prompt once and we don't touch it. No per-model tuning, no prompt tweaks mid-season, no coaxing a better answer out of one model than another. Changing the prompt would change the test, so the only thing that varies between models is the model.
Line-blind. The model never sees the betting line. It produces its own win probabilities from the data alone, so we're measuring forecasting skill, not an echo of the market.
One packet, no search. It gets the same point-in-time data (ratings, Statcast, bullpens, park/umpire, situational splits) ~3 hours before first pitch, with web search off. It's the model we're measuring.
Every market, graded on its own. From that one forecast we score the moneyline winner and the run total (over/under) on MLB, and the winner and method of victory on UFC — each against the final result, and the money markets against the closing price (ROI).
Graded in public, no do-overs. One set of calls per game and we live with it. See the full leaderboard and the Can AI beat Vegas? essay.
Compare DeepSeek with the field
GLM 5.2's graded record, line-blind.
Claude Opus 4.8's graded record, line-blind.
Grok 4.5's graded record, line-blind.
Claude Sonnet 5's graded record, line-blind.
GPT-5.6 Luna Pro's graded record, line-blind.
Gemini 3.6 Flash's graded record, line-blind.
Frequently asked
Can DeepSeek predict sports?
We grade DeepSeek (DeepSeek V4 Pro) in public. Across 486 MLB games forecast line-blind — no betting line, no web search — it is 273–213, a 56% winner accuracy, with a Brier score of 0.249, ranking 8 of 8 models on the board. Every call is locked before the game and graded against the final result.
What can DeepSeek predict — just winners?
No — DeepSeek makes a full forecast for every game, and each part is graded separately in public. On MLB it calls the moneyline winner and the run total (over/under). On UFC it calls the winner and the method of victory — KO/TKO, submission or decision. So far its over/under calls are 52% accurate over 468 games. Its UFC method calls are 55% accurate over 64 fights.
Is DeepSeek good at sports betting?
Every pick is also scored as ROI at the closing market price, so the number is honest about the vig. So far DeepSeek V4 Pro has returned +1.0% per unit over 486 graded MLB games. Beating the closing line is a high bar — this page tracks how close it gets, updated nightly.
How accurate is DeepSeek at predicting baseball?
Over 486 graded MLB games, DeepSeek V4 Pro has picked the winner 56% of the time, with a Brier score of 0.249 (0 is perfect, 0.25 is a coin flip). The 'pick the home team' baseline over the same games is 52%.
Track DeepSeek and the whole field.
Who's hot, who's slipping, where the AIs disagree — free in your inbox every morning.