Uber declares the tokenmaxxing era over on the same call it holds AI spend flat

On Wednesday Uber reported Q2 2026 alongside two coordinated messages about its AI bill. The CFO, Balaji Krishnamurthy, told analysts that per-token costs had fallen enough over the last several months to keep total AI spend "broadly stable" even as adoption rose. The CTO, Praveen Neppalli Naga, posted on X the same day that this was, according to Business Insider's account, another signal that the so-called "tokenmaxxing era" is ending. The interesting bit is that the company saying so is the company that publicly blew through its full-year Claude Code budget in four months.

What was actually said. Krishnamurthy's prepared remarks, as reported by Business Insider and summarised on Uber's earnings call, described four cost-management moves: setting better model defaults for different use cases, routing some tasks to lower-cost or open-weight models, giving employees visibility into their own spend, and, per Naga's post, prompt caching. Krishnamurthy also told analysts, per the same reporting, that Uber is seeing a doubling in code output per engineer, with what the CFO described as near-100% adoption of AI coding tools across engineering. Naga's framing, per India Today and Business Insider's account of the X post, was that the next phase will not be about who spends the most tokens but about who uses them most efficiently.

What actually happened earlier this year. Per Inc.'s reporting from June, Uber exhausted its 2026 Claude Code budget by April, roughly four months after rolling the tool out to about 5,000 engineers, with monthly API costs per engineer running $500 to $2,000. Inc. reported the company then imposed a $1,500 per-employee monthly cap on agentic coding tools including Cursor and Claude Code, with dashboards and an override request mechanism. That is the "before" picture for Wednesday's "after".

The inference-economics read

The frame that fits most cleanly is inference economics: the cost of running models, not training them, as the binding constraint. Uber is a large enterprise buyer publishing a data point most labs would rather narrate themselves. Per-token prices are falling, model tiering (routing easier work to cheaper models) is working, and prompt caching (re-using computed context across calls) is doing enough real work to keep an entire company's AI line item flat while the user base grows.

The labs' pricing curves imply the same thing from the demand side: falling unit prices absorbed by rising consumption, with total spend held rather than reduced. If a $500–$2,000-per-engineer-per-month run rate has been squeezed under a $1,500 cap for most engineers and total spend is still flat while adoption quadrupled, the arithmetic is doing something. Either the average is well below the cap, or the mix shift to cheaper and open-weight models is carrying more weight than the caching story admits. Uber has not disclosed the mix.

The FDE structural read

The second frame is where enterprise AI actually gets deployed. Uber's four-measure programme, caching, model defaults, per-engineer dashboards, open-weight trials, is an internal capability build, not a vendor conversation. Uber is not asking Anthropic to make Claude Code cheaper. It is internalising inference-cost discipline: pushing decisions about model choice and prompt structure down to the engineer, with visibility and a cap as the guardrails.

The shift is from a central AI centre of excellence buying seats to embedded engineers held accountable for their own token line. The vendor sells the model; the customer builds the routing, the caching, the dashboards, and the override queue. For Anthropic and its peers this is the customer telling them, politely, that the interesting margin is going to sit inside the customer, not the vendor.

What the productivity claim will and will not carry

The awkward part of Wednesday's story is that in May, COO Andrew Macdonald said on the Rapid Response podcast, per Business Insider, that the link between higher token spend and shipped consumer features was not yet there, and that the trade was getting harder to justify. Three months later the CFO tells analysts, per Business Insider's reporting of the call, that code output per engineer has doubled and adoption is near-universal.

Both claims can be true and still not close the gap Macdonald named. Doubled code output is a volume metric measured inside engineering. "Useful features shipped to users" is a business metric measured outside it. Uber has not, on the evidence disclosed this week, published anything that closes the distance between the two. The measurement gap has been reframed as a productivity win; whether it is one depends on numbers Uber has not put on the table.

What this is a case of

Set Uber alongside two other data points from the last few weeks. Per Business Insider's reporting, Microsoft cancelled Claude Code licences across its Experiences and Devices division before Uber's Q2 report. Per Business Insider's account of Naga's post, Coinbase is experimenting with model switching — routing harder tasks to frontier models, easier ones to cheaper alternatives. Both are the same move: enterprise buyers pulling back from frontier-only deployments and building their own routing layer.

For the agentic-coding SaaS layer, Claude Code, Cursor, Copilot on per-token pricing, the question is net dollar retention. If the biggest, most public Claude Code customer is telling the market it kept spend flat by moving work to cheaper and open-weight models, that is a pricing-power signal from the demand side. It does not say the product is failing. It says the pricing model is under pressure from customers who can afford to build the discipline in-house.

One thing I did not find. Uber has not disclosed the split between frontier and open-weight token consumption, nor the pre- and post-cap distribution of per-engineer spend. Without those numbers, "cost per token has declined" is a direction, not a magnitude, and the share of the flat-spend outcome attributable to caching versus model substitution is unknowable from outside.

Footnotes