Uber halved the cost of an AI session by routing work away from frontier models

Two of Uber's numbers do different work. A session now costs 48% of what it used to and a request costs 66% of what it used to ^[4][5]. Divide 0.48 by 0.66 and you get 0.73, so a session holds roughly 27% fewer paid requests than before, if the average request inside a session costs about what the average request costs overall ^[17]. Some of the saving is a cheaper unit. Some of it is fewer units per piece of work, which is what automatic compaction above 400K tokens and prompt caching inside Uber's own harness, Minions, are for ^[7].

The three companies are not evidenced the same way. Uber's percentages come from sources who spoke with Gergely Orosz ^[10], alongside Uber's own engineering blog post, which describes dozens of optimisations that cut token cost with no noticeable change in the quality of code its agents generated ^[8]. Pinterest's figure was said out loud on an earnings call ^[11]. AT&T's came through The Information ^[14]. Only AT&T put a number on the other side of the trade: a 2% decrease in output quality for a 56% cut in the bill ^[14], which is 28 points of cost for each point of measured quality ^[19].

"With open models, we are achieving cost per transaction at less than 8% of the cost of comparable closed proprietary models," Pinterest chief executive William Ready said ^[11]. Under 8% is a reduction of more than 12.5 times ^[18], at the far end of the 2-20x range Orosz gives for open models against frontier ones ^[9]. Ready credited compact models built for Pinterest use cases and open models post-trained on the company's own data inside its own secure cloud ^[12].

In May, Uber's COO said it was getting harder to justify spending on tools like Claude Code without seeing benefits from the leading models, in Orosz's account ^[3]. The list of tactics companies were trying at that point included per-developer monthly usage limits and knowledge-sharing sessions on cost-effective model use ^[1]. Neither appears in the account of how Uber got its number ^[20]. What is on the record is model selection, weekly benchmarking and request hygiene ^[7].

A skeptic would say this is arbitrage on a price gap that closes. Open models cost 2-20x less today ^[9], and nothing stops frontier vendors from repricing. The part of Uber's method that survives that is the cadence: benchmarks re-run every week against real work and updated ^[7]. That is also the part a twenty-engineer team cannot copy from a memo, because the same list assumes an in-house harness and an inference provider underneath it ^[7].

Austin, quoted by The Information on AT&T's results, said open source models are "just as good or better" than older models sold by the likes of Anthropic and OpenAI ^[15]. AT&T's software developers still rely on cutting-edge models for complex tasks, and the company employs 100,000 people ^[16][13].