Uber keeps its AI bill flat by choosing which model runs which workload

Seventy percent is an attribution statistic. It records which process opened the pull request ^[1], and Uber's own description of the workflow keeps people in place as reviewers and escalation targets while managed agents start a growing share of sessions ^[7]. Its uReview system performs AI code review on every pull request ^[8]. The binding resource in that arrangement is the attention of whoever handles escalations, plus the credibility of the machine reviewer standing between an agent's diff and the main branch.

The skill library deserves its own line in the org chart. Divide more than 30,000 daily skill executions by more than 3,600 built skills ^[2][3] and the average is roughly eight runs per skill per day ^[18]. The post does not publish the distribution, so we do not know how much of that traffic concentrates in a handful of skills, and nothing in the record lets us guess responsibly. Either way, 3,600 internal artifacts is a catalogue, and a catalogue needs owners, deprecation rules, and some way to distinguish a skill that earns its keep from one nobody has run since the week it was written.

The cost mechanism is the transferable part. Uber says the vendor sets the token price while Uber picks which model runs which workload ^[9], with the efficiency gains coming from routing internal workloads more intelligently inside standard tier pricing rather than from a bespoke discount ^[13]. The method is unglamorous: build a benchmark from the agent's real work, run it on a harness that serves any model, frontier or open-weight, behind one interface, then keep moving as the frontier changes every few weeks ^[10]. For uReview, changing models raised F1 while cutting cost per pull request ^[21]. The lever the post calls most impactful is duller still, and it is a default: subagents are pointed at a weaker, cheaper model because their tasks often do not need frontier reasoning ^[11][12].

Then the budgeting arithmetic. Weekly active users across the agentic tools grew sevenfold between February and August 2026, and weekly requests grew 9.4x ^[4][5]. Divide one by the other and requests per user rose by about a third across those six months ^[19][20]. Total spend has been relatively stable since April ^[6], so consumption per head climbed inside a fixed envelope. A forecast built on that pattern should use requests and the routing policy behind them as the denominator, not seats.

Two limits apply. This is a self-reported blog post from a company with a platform organization most engineering groups do not have, and no absolute spend figure appears in it. Uber concedes part of the first limit: it says its measured cost reductions are specific to its environment and that the portable element is the methodology of benchmarking real work ^[14]. It is also worth being precise that "relatively stabilized since April" is not a published cost per completed task, and because the growth curve is not split before and after April, the implied unit-cost improvement cannot be derived from what has been shared.

The trade-off to name this quarter is ownership. Standardizing on a single vendor's packaged agent buys speed and hands the default model choice to that vendor's release cycle, which is precisely the lever Uber credits with the most movement ^[12]. Building the harness and the benchmark set costs platform headcount now and buys the ability to switch when the Pareto frontier moves ^[10][15]. That is a tooling decision this quarter, and next quarter it determines whether you set your cost curve or simply receive one.