Home / Guides / Can LLMs predict sports results?

Can LLMs predict sports results?

It's early days, and that's exactly why we're watching. Most of these tools run several AI models side by side, ChatGPT among them, alongside Claude, Gemini, Grok and others, so this guide looks at large language models (LLMs) as a whole. We rate 7 of them today; 2 publish enough for a Performance score and none of the 2 are in profit at level stakes so far. Samples are still small, most records are self-reported, and nothing here is betting advice. We're tracking the movers from the start, every week and every month, so you can follow the same history we do. (Data as of 21 Nov 2026.)

Last updated 21 Nov 2026Built from our ratings data; entries are re-checked on our published scheduleReviewed weekly · last reviewed 17 Nov 2026How we score

Three ways these tools use LLMs

Using language models for sports tips is still in its infancy. None of the tools below is a single-brand ChatGPT product: most run several models together, and a few name one live model that can change over time. The services we rate are not all doing the same thing:

  • A plain chatbot tip. Asking ChatGPT, Claude or Gemini in a chat window can sound confident, but it usually has no timestamped ledger of its own tips, may miss today's team news or prices, and isn't bound by UK tipster ad rules. That alone is not something we can score.
  • Multi-model leaderboards and consensus. Sites such as ModelFights, Kickscore AI, BetBrothers AI, Football Arena, AI Betting Tips and ScoreGPT send the same brief to many LLMs, publish every (or most) picks, and rank models, or blend them into a consensus. That at least creates a public trail.
  • One live LLM wired into live odds. ProphetAI is the clearest example we rate: it reads bookmaker prices, posts each pick before the event with stake and reasoning, names which model wrote it, and settles a public virtual bankroll. The model on duty can change; the ledger is what we score.

A trained, sport-specific statistical model is a different product again (see Does AI sports betting work?). This page is only about tools that say they use large language models to pick or grade bets.

Which models each tool names

Line-ups move quickly. Providers swap GPT, Claude, Gemini and Grok versions within weeks. The table uses what each site showed on 5 Nov 2026, next to our reading of its public record. It is a snapshot, not a permanent roster.

LLM-based tools we rate: approach, models named and record
ToolApproachModels named on its own pages (checked 5 Nov 2026)Our reading of its public record
ProphetAI
Multi-sport
Single live LLM + live oddsNames the live model on its ledger (Claude Opus 5.5 when we last read it; earlier periods include GPT-5.6 Sol and other Claude versions)-0.70% over 465 level-stakes bets (self-reported, advised odds; Performance 50/100)
Kickscore AI
Multi-sport
Multi-model leaderboard + consensusAbout 17 models including ChatGPT, Claude, Gemini, Grok, DeepSeek, Mistral and Qwen-2.60% over 546 level-stakes bets (self-reported, advised odds; Performance 49/100)
ModelFights
Multi-sport
Multi-model scoreboardOpenAI GPT, Claude, Gemini, Grok, DeepSeek, Llama, Mistral and others on a public arenaStake and price basis not stated, and the published ROI column is not consistent with the units figures.
BetBrothers AI
Multi-sport
Multi-model with a value gateSix named models: OpenAI gpt-4.1-mini, Claude, Gemini, Grok, Mistral and DeepSeekRecords are per model and per sport; no single combined record is published. On the cricket leaderboard every model is behind (−4.9% to −15.3% at $100 per bet).
Football Arena
Football
Per-model football benchmarkOpenAI, Anthropic Claude, Google Gemini, DeepSeek and others, scored separatelyPublishes separate records for each AI model and provider, not one record for the service, so there is nothing single to score. Figures are quoted as stated.
AI Betting Tips (aibetting.tips)
Football
Multi-model football leaderboardFourteen language models including ChatGPT, Claude, Gemini, Grok, DeepSeek and QwenPublishes separate records for each AI model, not one record for the service, and doesn’t state the stake or odds basis on its public pages. Figures are quoted as stated.
ScoreGPT
Multi-sport
Consensus of several modelsFive-model consensus: GPT-6.1 Sol (OpenAI), Claude Opus 5.5, Grok 4.7, DeepSeek-V4-Pro and Kimi K3.1 ProPublishes hit rate only (1,010 graded picks for the top model, 4,955 across five models); no stakes, prices or returns.

Model names are quoted from each tool's own public pages on 5 Nov 2026; line-ups change, so check the linked entry. Figures use our usual Performance rules. Self-reported records are labelled. Past results do not predict future returns.

What the records show so far

Honesty first: of the 7 LLM tools we rate, only 2 currently publish a level-stakes record we can fully score, and none of the 2 of those show a profit. The scored records so far are self-reported and settled at the prices the tool itself logged (advised-odds style), which can be hard to match in practice.

  • ProphetAI: -0.70% over 465 bets (self-reported, advised odds), Performance 50/100.
  • Kickscore AI: -2.60% over 546 bets (self-reported, advised odds), Performance 49/100.

The other 5 are still building or only publish per-model tables, hit rates or units we can't treat as one level-stakes service record. That is normal at this stage of the market, and it is why a living tracker matters more than a one-off verdict.

UK advertising rules say tipsters must not imply success is guaranteed (ASA guidance). We apply the same care: no profit promises, every figure dated and labelled, 18+ only.

We're tracking this from the start

Because LLM tipping is young, the useful story is the path, not a single screenshot. AI Sports Ratings keeps a full historical record of which tools move, which models they name, and how their published figures change:

If you want the upbeat version: we get to watch a new class of tipsters grow up in public. Some will publish fuller ledgers; some will swap models; some will stall. We'll still be here with the same rules when the samples are measured in thousands instead of hundreds.

What we're watching next

  • Sample size. Records climbing past a few hundred settled bets, then past 1,000, with losers kept in view.
  • Price basis and CLV. Clear statements of advised odds vs starting price, and any closing-line value (see What is closing line value?).
  • Model versions. When a tool switches from GPT to Claude (or the reverse), whether the ledger resets, splits by generation, or quietly mixes eras.
  • Consensus vs single model. Whether multi-model agreement actually beats the best single model on the same fixtures, on their own numbers, checked by us.
  • Independently proofed trails. Third-party or downloadable bet-by-bet logs we can recompute, not only dashboard totals.

What would change our view

We'd get more constructive on any LLM tool that publishes a long, level-stakes ledger with a stated price basis, keeps every loser, timestamps picks before the off, and ideally lets us recompute the totals. We'd get more cautious if records shrink, filters appear after the fact, model eras are blended without a note, or claims drift from the ledger. Until then, treat every tip, chatbot or otherwise, as unverified entertainment, not a plan to make money.

LLM-based tools we rate (7)

Tools whose own description says they use large language models to pick or grade bets. Those with a full Performance score are ranked by Performance, then Transparency score; the rest are listed as not ranked yet. We leave realtime.tips out of this guide for now because its site currently shows an SSL certificate warning.

  1. 1

    ProphetAI

    -0.70% ROI over 465 level-stakes bets (26 Sep – 30 Oct 2026), self-reported. Price basis: Odds on offer when each tip was posted, as its record states.

    • Self-reported⚑ Advised odds, may be hard to matchMulti-sportRecord period: 26 Sep – 30 Oct 2026 (tips at a flat stake)Checked 20 Nov · next re-check 20 Dec
    50
    PerformanceBreak-even record
    55
    TransparencyMedium transparency
  2. 2

    Kickscore AI

    -2.60% ROI over 546 level-stakes bets (24 Sep – 6 Nov 2026), self-reported. Price basis: Odds recorded when the bet was named (CLV against close also shown).

    • Self-reported⚑ Advised odds, may be hard to matchMulti-sportRecord period: 24 Sep – 6 Nov 2026 (AI consensus, our published-prediction line)Checked 6 Nov · next re-check 6 Dec
    49
    PerformanceLosing record
    25
    TransparencyLow transparency

Not ranked yet (5)

Provisional (50 to 199 settled bets) or not enough published for a full Performance score. Listed by what they publish, not ranked.

  • ModelFights

    Stake and price basis not stated, and the published ROI column is not consistent with the units figures.

    Multi-sportChecked 7 Nov · next re-check 7 Dec
    Low data
    PerformanceInsufficient data
    75
    TransparencyHigh transparency
  • BetBrothers AI

    Records are per model and per sport; no single combined record is published. On the cricket leaderboard every model is behind (−4.9% to −15.3% at $100 per bet).

    Multi-sportChecked 6 Nov · next re-check 6 Dec
    Low data
    PerformanceInsufficient data
    35
    TransparencyMedium transparency
  • Football Arena

    Publishes separate records for each AI model and provider, not one record for the service, so there is nothing single to score. Figures are quoted as stated.

    FootballChecked 30 Oct · next re-check 29 Nov
    Low data
    PerformanceInsufficient data
    10
    TransparencyLow transparency
  • AI Betting Tips (aibetting.tips)

    Publishes separate records for each AI model, not one record for the service, and doesn’t state the stake or odds basis on its public pages. Figures are quoted as stated.

    FootballChecked 27 Oct · next re-check 18 Dec
    Low data
    PerformanceInsufficient data
    0
    TransparencyLow transparency
  • ScoreGPT

    Publishes hit rate only (1,010 graded picks for the top model, 4,955 across five models); no stakes, prices or returns.

    Multi-sportChecked 10 Nov · next re-check 10 Dec
    Low data
    PerformanceInsufficient data
    0
    TransparencyLow transparency

Questions and answers

Can ChatGPT predict sports results?

On its own, a general ChatGPT chat has no verified tip history we can score and may not have live team news or prices. Some of the tools above include an OpenAI GPT / ChatGPT line inside a wider LLM panel or ledger. Those are what we rate, not a private chat window.

Do any LLM tipping services make a profit?

Of the 2 LLM-based tools with a Performance score on 21 Nov 2026, 0 show a profit at level stakes. Today's scored records are self-reported. Each evidence type is labelled above; past results do not predict future returns.

How can I follow your tracking?

Start with the latest weekly tracker issue (week of 16 Nov 2026) · State of AI predictions: 2026-11 · full AI tipping tracker. New weekly issues land on Mondays; the monthly State of AI predictions summarises the field. Every tool page shows its last-checked date.

Is this betting advice?

No. Ratings are information for adults (18+). We never publish tips or take bets. If gambling is causing harm, see BeGambleAware.org or call 0808 8020 133.

How we check the records

The transparency checklist sets out what counts as evidence for each Transparency criterion, the same for every listing. The AI tipping transparency report 2026 shows how many of the AI tools and tipsters we rate meet each criterion.

18+ only. This is not betting advice. We rate published records; we never give tips or take bets. No tool or tipster can guarantee profit, past results don't predict future returns, and subscription fees are not included in any figure. Some links on this site are affiliate links; they never change a score or a rank, and scores are never for sale (disclosure). If gambling is causing harm, visit BeGambleAware.org or call the National Gambling Helpline on 0808 8020 133 (more help).

More guides and rankings

All guides →