What happened
On August 27, 2026, Uber distinguished engineer Uday Kiran Medisetty published a detailed account on Uber's own engineering blog of how the company runs AI coding agents at scale. The headline numbers: between February and August 2026, weekly active users across Uber's agentic AI tools grew 7 times over, and weekly agent requests grew 9.4 times over. Over that same window, total AI spend "relatively stabilized" starting in April. More than 70% of Uber's pull requests are now attributed to local or cloud agents, and the company runs more than 30,000 agent skill executions a day. To separate real efficiency gains from the effect of simply using a cheaper model, Uber held one model fixed and measured its own optimization work from February to July. Cost per 1,000 model requests fell almost 34% from its peak. Cost per session fell 52% from its June peak. Axios independently confirmed these figures the same day, adding an on-record quote from Uber CTO Praveen Neppalli describing the shift as the end of what he called the "tokenmaxxing era." CNBC TV18 independently confirmed the same usage, cost, and technical details. The specific levers Uber describes are concrete, not abstract. It decomposed total AI spend into six measurable terms multiplied together (users, sessions per user, turns per session, requests per turn, tokens per request, and price per token) and assigned optimization work to each. It routes tasks to the cheapest model that clears a quality bar for that task, benchmarked on real work, reserving the most capable model for decomposition and evaluation. It caps interactive sessions at 400,000 tokens even on models that can handle a million. It extended its prompt cache window from the default 5 minutes to 1 hour for interactive sessions, since engineers routinely leave sessions idle longer than 5 minutes, which was silently forcing expensive full-price context rebuilds. And it removed 50,000 to 70,000 tokens of tool-schema overhead that was being reloaded into every session, replacing it with on-demand tool search.
Why it matters for business owners
Most medium-size businesses will never run Uber's scale of AI infrastructure. But the pattern behind this story is one almost every business adopting AI agents or AI coding tools will eventually hit: usage grows faster than anyone planned for, and the instinct is to assume the bill has to grow at the same rate. That assumption is exactly what this account challenges. Uber is not claiming AI got cheaper on its own, and it is not claiming it simply spent less by using AI less. It is reporting that a specific set of engineering choices, mostly about how requests are routed, cached, and sized rather than which model is used, changed the relationship between usage and cost. For a business evaluating whether to expand AI use into more of the company, or trying to understand why an existing AI tool's bill keeps climbing, that distinction changes what the right next move is. The question is not only 'how do we use AI less.' It is also 'is our current setup wasting money on the way to a correct answer.'
What owners should not misunderstand
This is not evidence that AI costs are falling for everyone, or that any business will see the same 34% or 52% reduction by copying a checklist. Uber's own post is explicit that its measured savings are specific to its environment, codebase, team size, and workloads, and that results will vary. Provider prices, discount structures, and cache pricing also change over time, so treat the mechanisms here (route by task complexity, cap context, tune cache windows, cut unnecessary preloaded overhead) as the transferable part, not the exact percentages. It is also not a story about Uber deciding AI usage was too high and reining it in. Usage went up, sharply, and Uber's stated response was to make each unit of that usage cheaper rather than to slow adoption. That is a different move than the flat per-employee spending cap Uber itself put in place earlier in 2026 after burning through its annual AI budget in four months, which we covered in an earlier post. Both are legitimate tools. They solve different problems: a spending cap buys time and visibility when costs are already out of control. The engineering work described here is what makes it possible to keep growing usage afterward without the bill growing at the same pace. And this is specific to AI coding agents and similar high-volume, agentic workloads. A business using a single AI chat tool for occasional writing or research is not going to encounter the same cost dynamics, and does not need Uber's measurement infrastructure to manage its AI spend.
The operational lesson
A spending cap and a cost-efficiency program are not the same fix, and businesses often stop at the first one. A cap answers "how do we stop this from getting worse." It does not answer "why does this task cost what it costs," which is the question that determines whether a business can grow its AI use without its AI bill growing at the same rate. Uber's account is useful precisely because it shows the second question can be answered concretely, without frontier-lab resources. Most of the individual levers are structural rather than exotic: send routine work to a cheaper model and reserve the expensive one for what actually needs it, cap how much context a session is allowed to accumulate, make sure a cache setting matches how people actually pause and resume work, and stop reloading information into every request that most of those requests never use. None of that requires custom infrastructure to understand, even if replicating Uber's specific dashboards does. The other detail worth sitting with: Uber chose visibility over hard limits for most of this. A live cost counter and usage nudges, rather than blanket caps, let people see the cost of what they are doing and adjust their own behavior. That is a governance choice, not just a technical one, and it is one a business of any size can apply without building anything close to Uber's tooling.
What a serious business should do next
Before assuming a growing AI bill means you need to use AI less, break the bill down the way Uber did: how many people are using it, how often, how many exchanges per session, how much each exchange actually costs, and on which model. Most businesses can answer the first two questions from a subscription dashboard and have no idea about the rest. That gap is usually where the waste is hiding, not in the number of people using the tool. Check whether your highest-volume AI workflow is running on the most expensive available model by default, out of habit rather than necessity. If a cheaper model handles the routine version of that task with output your team accepts, that is a routing decision worth testing on a sample before assuming it will hurt quality. If your AI tools support prompt caching, check whether the default cache window matches how your team actually works. A cache tuned for constant back-to-back use will not help a team that pauses for 15 or 30 minutes between messages, and the fix is often a configuration change, not a new tool. If you cannot currently see who on your team is generating the bulk of your AI spend and on what, that visibility gap is worth closing before you decide whether the answer is a cap, a routing change, or leaving things as they are.
The Atlacis view
This is the kind of story worth reading past the headline number. The interesting part of Uber's account is not that usage grew 9 times over. It is that the company can explain, in specific and checkable terms, why its cost per unit of work went down at the same time, and that explanation holds up across three independent sources with no conflicting figures. Atlacis helps business owners do the smaller, earlier version of that same audit: understand what your AI workflows are actually costing per unit of output, where the routing, caching, or context decisions are quietly wasting money, and whether a spending cap or a structural fix is the right tool for where your business actually is right now.
The short version
- Uber's own engineering blog reported on August 27, 2026 that weekly AI agent requests grew 9.4 times over and weekly active users grew 7 times over between February and August 2026, while total AI spend stayed roughly flat since April. Axios and CNBC TV18 independently confirmed the same figures.
- With the AI model held constant, cost per 1,000 requests fell almost 34% from its peak and cost per session fell 52% from its June peak, measured February to July 2026.
- This is not a repeat of the earlier Uber story about burning through its annual AI budget in four months and capping spending at $1,500 per employee per month. That was a crisis response. This is the later, more advanced fix: making each unit of AI usage cheaper rather than limiting how much of it happens.
- The specific levers Uber describes are transferable in kind, not in exact percentage: route routine tasks to a cheaper model, cap how much context a session accumulates, tune prompt cache windows to match real usage patterns, and stop reloading unused information into every request.
- Uber chose visibility (a live cost counter and usage nudges) over hard caps for most of this work, a governance choice available to a business of any size, independent of its technical infrastructure.
- Before assuming a growing AI bill means using AI less, break it down by users, sessions, requests per session, and cost per request. That breakdown usually reveals where the actual waste is.
Where ATLACIS can help
Sources
- Uber Engineering Blog: Running a Software Factory Efficiently at Uber Scale (Uday Kiran Medisetty, August 27, 2026)
- Axios via Yahoo Finance: Exclusive: Uber cuts AI costs even as usage jumps (Madison Mills, August 27, 2026)
- CNBC TV18: Uber is using nearly 10 times more AI without spending more on it (August 27, 2026)