World first
SharpBench
Which AI actually knows betting? Two boards, one rule: every model answers the same questions and reads the same slips, scored the same way.
Settlement · odds · rules
SharpBench
SharpBench Settlement Index; higher is better
Under each bar: price per 1M tokens, input and output.
Sports-1 sits 2.8 points behind the leader at 7× less per output token — and is cheaper than every model above it.
SharpBench Betslip
Reading the slip
SharpBench Betslip is Infersia’s proprietary AI benchmark for reading, extracting, deciphering and returning betslip data in a structured format.
The problem. On the surface, one assumes that a simple betslip is easy to read and extract data from. In reality there are thousands of different permutations of how a betslip can be presented, with an equal number of bet types mixed in between — parlays, accumulators, same game multis, trifectas, exactas, bet builders and many more. SharpBench Betslip tests AI models on their ability to read, extract and return correct data for all betslips from all worldwide betting markets.
Reading real screenshots
SharpBench Betslip
SharpBench Betslip Index; higher is better
Under each bar: price per 1M tokens, input and output.
Sports-1 sits 1.4 points behind the leader at 4× less per output token — and is cheaper than every model above it.
Thirteen sports
Where the models still get it wrong
Sports-1 against the average of the frontier models that have completed the full benchmark. The team sports are close to solved. Horse racing is not — withdrawals, place terms and pool arithmetic remain the hardest thing in betting for a language model, and the gap between the best and the rest is widest exactly there. It is also where Sports-1 beats the field’s average — the bookmaker rulebook it carries is built for precisely those questions.
- Ice hockey96.899.3
- AFL97.699.1
- Baseball96.698.7
- Basketball97.898.5
- NFL96.098.4
- Combat sports96.498.3
- Rugby league97.298.3
- Soccer96.498.1
- Golf95.697.1
- Cricket92.896.1
- Formula 195.095.1
- Tennis88.693.8
- Horse racing89.085.1
Sports-1 Board average
What SharpBench measures
Not trivia. Every question has one right answer that a bookmaker’s settlement desk would agree with, and a model gets no credit for being nearly right.
Settlement conventions
Rule 4 deductions, dead heats, each-way place terms, pari-mutuel pools, exchange reduction factors — the conventions that decide what a bet actually pays, across UK, Australian, US and exchange markets.
Odds and bet maths
Decimal, fractional and American conversion, overround, accumulators and Lucky 15s, Asian handicap outcomes, staking — the arithmetic every settlement rests on, checked to the cent.
Reading real betslips
Screenshots from real bookmakers, read into one canonical JSON — every bet, leg, stake, odds, boost and status — scored field by field against reviewed ground truth, decoys included.
Built by Infersia
We serve the models worth building on, live sports data included — and our own Sports-1 and Sports-2 sit on both boards, scored the same way as everyone else. Have a model you think belongs here? We’ll run it.