Key Points

  1. Uber exhausted its 2026 AI budget within the first few months of the year, prompting CTO Praveen Neppalli Naga to impose cost controls on internal AI coding tools.
  2. Uber's AIFX wrapper now enforces four defaults — 400,000-token compaction, medium reasoning effort, one-hour prompt cache for orchestrators, and no MCP tool schemas at startup — which helped cut cost per session 52% from its June peak.
  3. The case shows that managing context, caching and tool schemas can deliver material AI savings without changing models, a lesson relevant to any enterprise scaling agentic workloads.

How Uber Spent Its 2026 AI Budget in a Few Months — and What It Changed

Uber spent its entire 2026 AI budget within the first few months of the year, according to CTO Praveen Neppalli Naga, who spoke with The Information in April. The overrun was driven in part by a surge in agentic workloads: weekly agentic requests across Uber's offerings grew 9.4x between February and August, according to the company's August engineering post. Rather than simply cutting usage, Uber rebuilt the defaults that govern how its engineers' AI coding tools behave.

When an Uber engineer opens Claude Code today, the session automatically compacts at 400,000 tokens — even on models that support million-token context windows. Reasoning effort defaults to medium rather than high. The prompt cache for the main agent lives one hour, while sub-agents get five minutes. And no MCP tool definitions load at startup; instead, agents invoke individual commands through a central gateway as needed. Engineers did not choose these settings. Uber did, through an internal wrapper called AIFX that installs and configures the tools and A/B tests settings across employee cohorts.

The results, reported in aggregate, are substantial. Holding the model constant to isolate the effect of the optimizations, Uber says cost per 1,000 requests fell almost 34% from its peak, and cost per session fell 52% from its June peak. Overall AI spending has remained roughly flat since April even as agentic requests kept climbing, according to Uday Kiran Medisetty, a distinguished engineer at Uber, speaking at VentureBeat's Agentic Infrastructure Stack Summit on Oct. 28.

Uber is now pushing more work away from interactive sessions — which Medisetty called the least efficient way to run an agent — and into managed agents where the company can select models per task. It has also built a context graph of roughly 200 million nodes and edges that gives agents a map of Uber's internal systems rather than preloading all relevant documentation. In one internal example, the same model with the same prompt identified a data table in 38 seconds with the graph, versus 20 minutes and 9 seconds without it, when it incorrectly concluded the data could not be queried.

At a Glance

Main CompanyUber
Subject of the cost-control program and the source of the reported figures.
Key ExecutiveUday Kiran Medisetty, Distinguished Engineer, Uber
Presented Uber's AI cost strategy at VentureBeat's Agentic Infrastructure Stack Summit on Oct. 28.
Key ExecutivePraveen Neppalli Naga, CTO, Uber
Told The Information in April that Uber had used up its 2026 AI budget early.
Key TechnologyAIFX
Uber's internal wrapper that installs, configures and A/B tests settings for interactive coding tools.
Key TermAutomatic compaction at 400,000 tokens
Uber's default context limit before the agent condenses earlier conversation into a summary.
Key TermPrompt cache: 1 hour main agent, 5 minutes sub-agents
Cache lifetimes tuned to how long each agent type typically idles between requests.
Key Financial FigureCost per session down 52% from June peak
Uber's reported result across all optimizations, holding the model constant.
Key Financial FigureCost per 1,000 requests down almost 34% from peak
Second headline metric from Uber's August engineering post.
Key Financial Figure90% cost cut per stale-feature-flag cleanup
Result of benchmarking models and picking the right one for the task.
Key DateOctober 28, 2026
VentureBeat Agentic Infrastructure Stack Summit where Uber disclosed the details.

Why Uber's Four Defaults Matter More Than Any Single Model Choice

Why Uber's Four Defaults Beat Any Single Model Upgrade

Uber's cost problem was structural, not model-specific. Every agent request sends the full context back to the model, so a larger context makes each step more expensive. Past 700,000 or 800,000 tokens, Medisetty said, cache misses and per-request input tokens "cost a lot." The 400,000-token compaction threshold is a deliberate trade: enough context to keep sessions useful, not so much that every step pays a premium. Similarly, loading 100-plus MCP tool descriptions at startup would consume 50,000 to 70,000 tokens before an engineer types a word — a fixed tax on every session that Uber eliminated by routing tool calls through command-line commands instead.

Where the Savings Actually Come From

Uber did not break out savings per default, but the aggregate numbers are telling. Cost per session fell 52% from June to the October summit, and cost per 1,000 requests fell nearly 34% from peak. The single largest disclosed saving came from model selection: benchmarking models and picking the right one cut the cost per stale-feature-flag cleanup by 90%. That suggests the biggest lever is matching model capability to task, not squeezing context alone. Uber's August post also notes that sub-agents default to cheaper models for tasks that do not require frontier-level reasoning — a pattern any enterprise can copy.

Why T-Mobile's Failure and Adobe's Success Both Validate Uber's Approach

Context is not universally good or bad; it depends on the task. T-Mobile CIO Jeff Simon said at the same summit that loading all customer support documentation into one agent "didn't work at all" — the agent got confused by the volume. T-Mobile now keeps context narrow and routes questions to specialized agents. Adobe's Anjul Bhambhri, by contrast, said marketing tasks often need more context: knowing which offers ran over the past three years revealed that free shipping converted better among loyal customers than a 20% Black Friday discount. Uber's context graph is a middle path — give agents a map of the company, not the whole library, and let them retrieve what they need. T-Mobile tried a similar graph and dropped it because live customer calls do not allow time for graph traversal. The lesson is that context strategy must match latency requirements and task structure, not follow a single best practice.

What Uber Is Really Measuring — and What It Is Not

Uber tracks throughput (design documents, pull requests, deployments) and guardrails (incidents per code change, revert rates, review and cycle times). It also monitors whether engineers act on AI reviewer comments, aiming to post only high-signal feedback. Trust is built gradually: self-healing CI fixes were first sent to engineers in Slack for approval, then auto-accepted once acceptance rates proved out. But Medisetty gave no figures for any of these measures and did not provide a calculation of Uber's overall financial return on AI spending. That places Uber in the majority: 44% of respondents to VentureBeat's October VB Pulse survey judge AI spending by operational results, not defensible financial returns. For context, 38% of 156 respondents said they had delayed, scaled back or canceled an AI project in the past year because it cost more than expected — a reminder that Uber's discipline is still the exception, not the rule.

The Four Questions Every Platform Team Should Ask About Its Coding Agents

The most important insight: Uber cut AI coding costs by more than half without changing models — by controlling context, caching and tool schemas at the platform level.

For platform and engineering leaders running agentic coding tools, Uber's experience suggests four concrete questions to ask about your own setup:

  • What is your context limit, and when does automatic compaction trigger? Uber chose 400,000 tokens because cache misses and per-request input tokens become expensive past 700,000–800,000. If your tools default to a million-token window with no compaction, you are likely paying for context engineers never use.
  • What reasoning effort runs by default? Uber set effort to medium after finding that recent models perform better at medium and show "some degradation at high." High effort is not automatically better — it is more tokens for potentially worse output.
  • How long does the prompt cache live for each agent type? Uber uses one hour for orchestrators (which engineers step away from) and five minutes for sub-agents (which finish quickly). Mismatched cache lifetimes mean paying full price to reload context that should have been cached.
  • How many tokens of tool schema load before the first prompt? Uber's 100-plus MCP tools would have consumed 50,000–70,000 tokens at startup. Routing tool calls through a gateway instead eliminated that fixed cost.

Two additional levers worth testing: benchmark models per task — Uber cut the cost per stale-feature-flag cleanup by 90% just by picking the right model — and default sub-agents to cheaper models for tasks that do not need frontier reasoning.

Metrics to watch: cost per session and cost per 1,000 requests (Uber's fell 52% and 34% respectively from peak), cache hit rate, and throughput measures like pull requests and deployments per engineer. Uber has only recently begun seeing a "real trend line" in features shipped per engineer, so expect a lag between cost optimization and measurable productivity gains.

Risk & Opportunity Assessment

Commercial RiskMediumUber's 2026 AI budget was exhausted early, forcing cost controls; 38% of surveyed enterprises have delayed or canceled AI projects for the same reason.
Competitive RiskLowUber's optimizations are internal; competitors face the same cost pressures but Uber's disclosed 52% cost-per-session reduction gives it a temporary efficiency edge.
Regulatory RiskLowThe story involves internal engineering cost management with no disclosed regulatory exposure.
Reputation RiskLowUber is presenting itself as disciplined rather than wasteful; the disclosure of an early budget overrun is framed as a learning experience.
Technology DisruptionHighUber's shift from interactive sessions to managed agents with per-task model selection could reshape how enterprises deploy coding agents.
Commercial OpportunityHighUber's context graph (200 million nodes and edges) and AIFX wrapper represent reusable patterns that cut cost per session 52% and cleanup costs 90%.