Skip to content
The AI-First Web

Uber treated the settings behind its AI bill as a cost lever

Agent requests grew 9.4x from February to August; Uber says total AI spend has relatively stabilized since April

W
WebPulse Newsroom
AI-assisted · 4 min read
Share on X LinkedIn
Uber treated the settings behind its AI bill as a cost lever
In brief
  • Uber says weekly agent requests grew 9.4x from February to August 2026, and that total AI spend has relatively stabilized since April.
  • Uber credits optimizations across the board. It calls its cheaper default model for subagents the most impactful lever, and it does not say how much each lever saved.
  • Leaders should ask who sets the defaults in their AI tools and whether anyone tracks cost per completed task.

Defaults engineers never have to choose are a cost lever

It is easy to treat the AI bill as a price problem: the vendor sets a rate and the customer pays it. Uber's engineering team describes a different view. The vendor sets the token price, it writes, but Uber chooses which model runs which workload.

The lesson here is that the settings behind an AI tool are decisions, and decisions can be managed. VentureBeat reports that Uber's engineers did not pick these settings themselves. The company set them, and it tests them with A/B comparisons across cohorts of employees.

What Uber reported

Uber says AI tools now reach every phase of its software work. Agents running locally or in the cloud are credited with more than seven in ten of its pull requests. A pull request is a proposed code change that awaits review.

From February to August 2026, Uber says, weekly active users across its agentic tools grew 7x. Weekly agentic requests grew 9.4x over the same months. Separately, Uber says its total AI spend has "relatively stabilized" since April, which it attributes to "optimizations across the board."

The early months looked different. VentureBeat says Uber's CTO, Praveen Neppalli Naga, told The Information in April that the company had used up its 2026 AI budget within the first few months of the year.

The lever Uber ranks first: which model does the work

In interactive sessions, Uber says, token prices are fixed. What the company can shape is which model handles which tokens. Two defaults govern that: the model a session starts with, and the model that subagents use.

Uber calls the subagent default "the most impactful lever." Subagents carry out narrow tasks with set inputs, which often do not need a frontier model. So Uber defaults them to a weaker, cheaper model, and engineers can still override it. The main model plans the work and checks the results.

Uber also picks models by benchmark. It defines the most efficient choice as the one balancing cost per completed task, output quality and reliability. For its code-review agent, uReview, Uber says a model switch improved the F1 score, a measure of accuracy, while cutting cost per pull request.

Why every extra word is paid for again

An AI agent does not remember a conversation. On each turn it sends the model the full history, the project context and the results of tools it has used. Uber says anything that shrinks that payload adds up across a session.

Uber also notes that output, including the model's internal reasoning, is billed at multiples of the input rate on primary models. Cost therefore depends on how much the agent carries and how long it thinks.

Four more defaults, among several levers

Uber sets four further defaults for interactive tools. It does not say how much each one saved, so treat them as parts of a larger effort, not as the whole explanation.

First, compaction starts automatically at 400,000 tokens, even for models that accept a million. Compaction swaps the long earlier exchange for a short summary. Distinguished engineer Uday Kiran Medisetty told a VentureBeat summit that beyond 700,000 or 800,000 tokens, cache misses and input tokens "cost a lot."

Second, reasoning effort defaults to medium. Medisetty said recent models perform much better at medium and show "some degradation at high." He did not name the models or tasks behind that comparison.

Third, cache timing. Uber says cached text is read back at 0.1x the standard input rate. Writing it costs 1.25x for five minutes of storage and 2x for one hour. VentureBeat reports that the main agent's cache is kept for one hour and subagents' for five minutes. Uber's post says engineers often leave the main agent idle for over five minutes, so a short cache would expire and force a full-price reload. Subagents finish quickly, so five minutes is enough for them.

Fourth, tool definitions. Uber says that with more than 100 tools installed, their descriptions added roughly 50,000 to 70,000 tokens to the opening prompt. Those tokens were re-sent on every turn. Uber routes more than 1,000 internal and third-party servers through one gateway and lets agents call tools as command-line commands. It reports that this "code mode" cut tokens by more than 50% on small queries and more than 90% on bulk workflows.

Visibility instead of hard caps

Uber says it avoided strict spending caps. Its harness shows a live cost counter in the terminal. Slack alerts arrive at 50%, 80% and 100% of expected spend, and managers can approve higher tiers quickly.

A session dashboard flags 16 anti-patterns, each paired with its financial impact and a suggested fix. One example is a simple session run on a premium model when a cheaper one would do. Uber says the aim is for engineers to judge a task's return on cost themselves.

What the evidence does and does not show

These are Uber's own accounts, from an engineering organisation large enough to build its own gateway and a context graph of 24 million nodes. Uber says its measured cost reductions are unique to its environment and results will vary.

The sources give no total dollar figure for spend or savings. The transferable part is a method: benchmark real work, then choose the model that balances cost, quality and reliability.

Questions to put to your team

Who sets the default model, subagent model, reasoning level and cache timing in our coding tools, and who reviews them? Do we know the cost per completed task, not just the monthly invoice? How many tool descriptions load before an engineer types a word?

Do people see spend while they work, or only after the invoice arrives? When a default changes, does anyone check that quality held?

Tokens your agent does not send are tokens you do not pay for.

Produced by the WebPulse Newsroom with AI assistance from the original reporting credited below, and checked against that source by our editorial review. How we use AI.
Original reporting: Uber.