Any model with publicly known benchmark scores and inference costs goes, not just OpenAI's o series.
I will consider a model to be "o3-equivalent or better" if it scores ≥25% on FrontierMath (o3 scored 25.2%) and performs similarly on other benchmarks.
(Note that o3's exact inference costs in the configuration used for benchmarking are currently unknown IIUC, though this market description will be updated with exact figures if they become public. This market can still resolve even without exact figures if e.g. OpenAI announce an o4 that's "10x cheaper" for roughly the same performance.)
1,000People are also trading
this may be hard to resolve because the inference costs for specific benchmark performances or tasks can vary so much.
@JoshYou as a concrete example, let's say o4 costs the same per-token (for simplicity) and can achieve 25% on FrontierMath with 1/10 as many tokens as o3 did, but requires 1/5 as many tokens to match o3 on ARC-AGI.
What's worse, those ratios probably vary a lot depending on the performance thresholds with a given benchmark. For example, it's over 100x more expensive to get 88% on ARC-AGI with o3 than it is to get 76% on ARC-AGI with o3. So it could turn out that o4 is 5x cheaper than o3 at the 76% threshold, but over 100x cheaper at the 88% threshold.
@JoshYou Hmmm... Yeah, there might be a relatively high chance of this resolving N/A when you put it that way, but I'll do what I can when the time comes.