International Edition
Latest News
Technology

Uber Caps AI Coding Costs With Token Limits, Reasoning Defaults

Uber applied settings to control costs after using up its AI budget early in the year, according to CTO Praveen Neppalli Naga. When an engineer opens Claude Code at Uber, automatic compaction triggers at 400,000 tokens. Reasoning effort…

Uber Caps AI Coding Costs With Token Limits, Reasoning Defaults

Uber applied settings to control costs after using up its AI budget early in the year, according to CTO Praveen Neppalli Naga.

When an engineer opens Claude Code at Uber, automatic compaction triggers at 400,000 tokens. Reasoning effort defaults to medium. No Model Context Protocol tool definitions load into the active context at startup. Those configuration choices belong to Uber, not the individual developer.

The company instituted those guardrails following unexpected expenses that consumed its AI budget during the first few months of 2026, according to Praveen Neppalli Naga, Uber’s chief technology officer. Financial pressures affected external organizations as well. In an October infrastructure survey of 156 qualified respondents by Venturebeat, 38% reported delaying, scaling back, or canceling a planned project because it exceeded financial projections.

Weekly agentic requests across Uber climbed 9.4 times between February and August. Despite that surge, overall artificial intelligence spending remained flat after April, Uday Kiran Medisetty, a distinguished engineer at Uber, told attendees at the Agentic Infrastructure Stack Summit on Oct. 28.

Four Managed Settings

Medisetty detailed four ways Uber restricts interactive session expenses using A/B testing across employee cohorts. Interactive coding remains the least efficient method for running an agent, prompting the firm to shift workloads toward managed agents where it controls model selection.

Automatic compaction summarizes conversations past 400,000 tokens. Beyond 700,000 or 800,000 tokens, cache misses and input costs rise sharply, according to Medisetty. Reasoning effort defaults to medium because newer models show performance degradation at high effort levels.

Prompt caches remain active for one hour for orchestrator agents and five minutes for sub-agents. Uber also avoids preloading tool definitions from its more than 100 internal servers. Direct integrations would otherwise consume 50,000 to 70,000 tokens before an engineer types a single word.

Uber AI Optimization Metrics
Metric Type Reported Change Reference Point
Cost per 1,000 requests Fell nearly 34% From peak usage
Cost per session Dropped 52% From June peak
Feature flag cleanup cost Cut 90% Through model benchmarking

Mapping Internal Systems

To prevent agents from drowning in documentation, Uber developed a context graph representing its internal architecture. The graph expanded from 104 million nodes and edges in August to nearly 200 million nodes and edges by Oct. 28.

“What we give the agents is a mental map of how Uber works,” Medisetty told the summit audience. Agents receive the basic schema and traverse the graph independently to retrieve necessary source code.

Other corporations adopted different approaches to context management. “We said, 'Oh, we've finally got a big context window model. Let's load all of our customer support documentation into that agent,'” Jeff Simon, chief information officer at T-Mobile, said at the summit. T-Mobile subsequently shifted to narrow contexts and specialized routing agents.

“We don't have a lot of time,” Simon told the summit. T-Mobile now relies on a key-value database cache to guide agents based on common customer phrasing.

Operational Return Metrics

“maximize the throughput while keeping our quality guardrails the same.” Medisetty told the summit audience. Medisetty tracks design documents, pull requests, deployments, review cycles, and code reverts.

Automated code reviewers post targeted comments to reduce review fatigue. Continuous integration pipelines attempt automated fixes for failed builds, moving from manual Slack approvals to auto-acceptance once acceptance rates improve.

Financial accountability varies across the industry. Among respondents to Venturebeat’s October survey, 26% measured financial return defensible to a finance team, while 44% relied on operational results like time saved or volume handled.

Uber Technologies, Inc., founded in 2009 and launched publicly in San Francisco in 2011, grew from a luxury black-car service into the world’s largest ridesharing platform.

It now operates in roughly 70 countries and 15,000 cities, coordinating an average of 42 million trips and deliveries per day and surpassing 72 billion total trips since inception. In December 2016, the company acquired Geometric Intelligence to form Uber AI, a dedicated artificial intelligence research division that was later shut down in May 2020.

“Since we can do A/B testing with managed settings, it allows us to experiment on these settings so that engineers don’t have to really think about all of this, and we give the same defaults for them,” Medisetty told the summit audience on Oct. 28.

About the author: Technology Desk

Technology Desk at Archynewsy.