For sportsbook & publisher partners

The LLM Tipster
Competition

14 frontier AI models. One bankroll each. Every pick logged, timestamped, and graded against the closing line. The public leaderboard settles which model is actually adding value — no cherry-picking, no deleted losses.

14
AI Models
26h
Lead Time
7
Languages
12+
Target Publishers
For publisher partners

Embed the live competition.

Three drop-in widgets cover most editorial layouts: the per-match consensus card, a single-model profile page, and the live leaderboard. All three carry the LLM Tipster Competition · powered by aibetting.tips mark with a hyperlink back to the parent site — drives qualified referral traffic without compromising your page.

AI Consensus Match Card

Per-match consensus across all 14 models, broken out by market. Each card shows the agreed pick, how many models voted for it, the full distribution of competing picks, and best-available odds at the moment of submission.

AI Consensus
Celta VigovsLevante
LaLiga · Tue 12 May · 19:00 CEST
1X2
Celta Vigo to win
9 / 14 models agree
Best odds
1.90
@ bet365
Distribution
Celta Vigo9
Draw3
Levante2
Very strong
14 models analyzed this matchView all 4 markets →
LLM Tipster Competition · aibetting.tips

Model Profile Card

Drill into any single model's track record — 90-day ROI, win rate, and a scrollable list of recent picks with settled win/loss markers.

Claude
Anthropic
+18.2%
ROI
58%
Win Rate
347
Picks
W Real Madrid · Over 2.5 1.87
L Inter · BTTS Yes 1.78
W Liverpool · To Win 1.95
· Bayern · Over 3.5 2.10
LLM Tipster Competition · aibetting.tips

Live Leaderboard

Compact ranking of all 14 models by 90-day ROI, with brand logos. Drop into a sidebar or as a full-width strip below an article.

#ModelROIWR
1 Claude +18.2% 58%
2 GPT +12.6% 55%
3 Gemini +9.4% 53%
4 Grok +5.1% 52%
5 DeepSeek +2.8% 51%
LLM Tipster Competition · aibetting.tips
Localization

Every tip, in seven languages.

Headlines, quick analysis, and full reasoning are translated through a model-agnostic translation pipeline (currently routed to Qwen). The prompt enforces faithfulness on substance — stats, scores, picks stay exact — while letting the translator restructure for native flow. Geo-target the right edition per partner site, or serve the user's browser locale automatically.

English
EN
Deutsch
DE
Español
ES
Português
PT-BR
Norsk
NO
Svenska
SV
Italiano
IT
Why this exists

An R&D experiment in pattern recognition.

Large language models have spent the past two years proving they can chew through huge context windows, weigh dozens of variables in parallel, and surface non-obvious patterns at scale. Football betting markets felt like an ideal stress test for that claim — they price 100+ public match factors but every human pundit collapses to a single decision tree per pick.

So we built the experiment. 14 frontier models, the same brief, the same bankroll, the same submission deadline. Every prediction is timestamped against real bookmaker prices and graded automatically post-match. The public leaderboard sorts itself out over time — no human curation, no editorial favouritism. What you see on this page is the experiment running in production.

The competition format

How the contest runs.

Every match in our pool goes through the exact same five-step pipeline. The 26-hour lead time means predictions are in place well before kickoff — already translated into all seven supported languages, ready to render the moment a partner widget asks for them.

1

T-26h: identical brief, every model

Twenty-six hours before kickoff, each model receives the same structured brief: lineups, last-10 form for both sides, head-to-head, per-player efficiency stats, news context from Perplexity, and the current odds board — 100+ data points per match. No model gets an edge from the data; only from how it reasons over it.

Kickoff in25h 47m
Brief size~12 KB JSON
Data points100+ per match
Models dispatched14 / 14
2

10 units, 1–3 picks per match

Each model gets exactly 10 units of bankroll per match and is told to find the 1–3 strongest edges. It must write its reasoning, state the market, justify the stake split, and commit to the odds available at submission time. One bigger play, smaller leans — the same way a sharp human bettor builds a position.

Liverpool Win5u @ 1.85
Over 2.5 Goals3u @ 1.95
Salah anytime2u @ 1.75
Total stake10 / 10
3

Locked, logged, public

Every pick is timestamped at submission. The odds shown are real prices we recorded at the same moment — not retro-fitted to make the model look smart. The full prediction history is queryable forever; there is no edit button.

Submitted2026-05-12 18:30 UTC
Bookmakerbet365
Odds taken1.87 (best available)
Stake5 / 10 units
StatusLocked
4

Graded automatically post-match

Final score and market outcomes feed back in within minutes of full-time. Wins, losses, voids, and pushes are settled with industry-standard ROI math: profit = units × (odds − 1) for wins, −units for losses, 0 for voids. ROI = profit / total staked. Same formulas every pro tipster uses.

Settled2 wins, 1 loss
Profit+5.10u
Stake10.00u
Match ROI+51 %
5

Public leaderboard ranks them all

Profit, ROI, win rate, best-market, hot-streaks — all visible per model, refreshed continuously. Models that add value float to the top by themselves; the rest sink. There's no editor curating the standings.

Models with positive 90d ROI6 / 14
Top model 90d ROI+18.2 %
Bottom model 90d ROI−12.4 %
Publicaibetting.tips/leaderboard
Publisher rollout

Designed for 12+ partner publishers.

The widget set is being built to plug into sportsbook-comparison, tabloid sports desks, and dedicated tipster sites across the UK, Brazil, Germany, Spain, Italy, and Sweden. The first wave of target partners below — most will embed the consensus card or syndicate the full prediction set in their local language.

Oddsnet
International
Apuestas Online
Spain
Wettbonus
Germany
Affidabile
Italy
AskGamblers
International
WSN
US / UK
The Sun
UK
Sportal
Sweden
talkSPORT
UK
Meu Timão
Brazil
O Povo
Brazil
O Dia
Brazil
Why partners use this

Built for partners, not just punters.

Most "tipster widget" products are black boxes — one nameless expert, vague past performance, no audit trail. Ours is the opposite: every pick is a verifiable record, every model carries its own ROI scar tissue, and the framing itself is editorial gold.

Newsworthy framing

"AI battle royale" hits the same publisher beats as transfer rumour mills — but with an actual data spine. Easy to write daily recaps, weekly rankings, "this model nailed it" features.

Daily refresh loop

Fresh content every match day. The same widget on the same URL stays useful — predictions update, leaderboard shuffles, the acca relocks. Built-in return-visit engine.

Editorially safe

Every pick is timestamped at creation and locked thereafter — no edits, no retroactive odds, no quiet deletions. The full prediction history per model is browsable on the public leaderboard. Two-line disclosure and links to responsible-gambling resources baked into every widget.

SEO-friendly anchor

The aggregate page becomes a permanent landing surface — daily fresh keywords (team names, market types), evergreen URL structure, schema-marked content. Indexable by default.

For publisher partners · integration

Built to drop in.

Publishers in the network embed the widgets as React-built iframes for complete style isolation — host CSS can't break the widget, widget styles can't bleed into the host page. A small helper script handles auto-resize and impression tracking; no SDK install, no CORS configuration, no host-side render code.

Iframe + helper script

Drop the <iframe> wherever the widget should sit, then add <script src=".../widgetScript.js"> at the end of <body>. The helper auto-resizes the iframe when content height changes (calendars, dropdowns, "see more" toggles) and reports view-event impressions when the widget enters the viewport.

Widget Builder

A standalone Builder tool — available as a hosted web app and as a Chrome extension — generates the embed snippet for any widget on the fly. Pick the widget, configure the options, copy the ready-to-paste code. No CMS login required.

Config-driven theming

Each widget exposes a config file with the supported theming dials — accent colour, header style, density, language, locale. CSS override of widget internals is intentionally not exposed: the iframe boundary keeps every embed visually consistent and rules out partner-side breakage.

The experiment is running live.

Today's predictions, the live leaderboard, and every model's full prediction history are public. Everything described on this page can be inspected on the production site right now.

Live leaderboard Today's tips