When multiple AI model providers compete on price and performance, their pricing doesn't perfectly track their capabilities. One provider might charge more per token for a model that's worse at math but better at code. Another might have a cheaper model that outperforms on a specific benchmark.
Note (2026-11-15; read against the paper's full text): the paper's empirical result is one case study, SWE-bench GitHub issue resolution with two models, GPT-5 mini and DeepSeek v3.2, where simple arbitrage strategies yield net profit margins of up to 40 percent; the essay's list of verifiable domains (math, code and factual queries) is the essay's own (no hit for math or factual queries in that role). The paper adds that robust policies fitted on one part of SWE-bench (it splits Django issues from the rest) remain profitable, that the 95 percent confidence intervals are wide because of small samples, that a slightly larger search budget of about 10 dollars allows consistently fitting a profitable policy, that distillation creates strong arbitrage opportunities potentially at the expense of the teacher model's revenue, and that multiple competing arbitrageurs drive consumer prices down while easing market entry for smaller providers; the essay mentions none of these.
The authors (arXiv:2603.22404) formalize this as computational arbitrage. In domains where outputs are verifiable — math, code, factual queries — you can route tasks to whichever provider offers the best price-performance ratio, accept only correct answers, and profit from the pricing inefficiency. The net margins reach up to 40% on verifiable domains.
The strategy is simple: given a task, estimate which provider will solve it correctly at the lowest cost, submit, verify, and pocket the difference between what the customer pays and what the cheapest correct solution costs. The verification is key — without it, you can't distinguish correct from incorrect outputs, and the arbitrage collapses.
The through-claim: AI model markets have the same structure as financial markets — when multiple providers price overlapping capabilities differently, arbitrage opportunities exist. But unlike financial arbitrage, computational arbitrage requires verification. The margin depends on your ability to check answers, which means it exists only in verifiable domains. Unverifiable tasks have no arbitrage because you can't tell if you got a good deal.