Evidence correction, 14 October 2026: earlier versions of this post misdated Uber's Claude Code rollout and The Information's report, mixed two different Uber adoption metrics, and misstated what DX, Deloitte, Augment Code, GitHub and AI.cc reported. They also attributed an arXiv study to Microsoft Research and cited developer surveys and token statistics we could not find. Those claims are removed or corrected below.
Uber's AI budget lasted four months.
In December 2025, Uber opened Claude Code to its engineering organization, roughly 5,000 engineers according to Forbes. Adoption moved fast. By March 2026, 84% of Uber developers were agentic coding users, and Claude Code use had grown from 32% of engineers in December to 63% in February. Nearly 95% of Uber engineers used AI tools every month. Around 70% of committed code came from those systems.
Then the bill arrived. Uber's CTO Praveen Neppalli Naga told The Information that the company had to revisit its financial assumptions after spending exceeded projections much earlier than expected. The entire 2026 AI budget was exhausted by April.
Per-developer costs ranged from $150 to $250 per month on average, but heavy users ran $500 to $2,000 per month. Startup Fortune argued that finance teams had no playbook for token-based billing at that scale.
Source: The Information, "Uber CTO Shows How Claude Code Can Blow Up AI Budgets," April 2026
Source: Forbes, "Uber Burns Its 2026 AI Budget In Four Months On Claude Code," May 17, 2026
Source: The Pragmatic Engineer, "How Uber uses AI for development: inside look," March 2026
Source: Startup Fortune, "Uber Burned Its Entire 2026 AI Budget in Four Months," May 2026
This is not just an Uber problem.
The Pragmatic Engineer newsletter has been tracking the same pattern across multiple companies. At one fintech, token spend rose about 10x in six months. A seed-stage startup went from $200 to $3,000 per developer per month across its seven developers. Some individual developers are spending $500 a day on Claude Code alone.
Source: The Pragmatic Engineer, "The Pulse: token spend breaks budgets – what next?", April 30, 2026
DX's survey of engineering leaders tells the planning side of the story. When asked about their 2025 spending, 38.4% of leaders reported spending $101 to $500 per developer per year on AI tools. Only 10.5% were spending over $1,000. For 2026, many planned to allocate 1 to 3% of their total engineering budgets.
DX advises that companies giving AI tools to every developer should expect to spend $500 per developer per year at the low end and $3,000 or more at the high end. It notes that almost all companies take a multi-vendor approach, one reason it expects AI costs to keep rising in 2026. Compare that with Uber, where the average engineer cost $150 to $250 a month, or $1,800 to $3,000 a year.
Source: DX, "How are engineering leaders approaching 2026 AI tooling budgets?", October 2025
Deloitte's January 2026 analysis, which it calls "the pivot to tokenomics," confirmed the trend at the enterprise level. AI has become the single fastest-growing line item in corporate technology budgets, consuming a quarter to one-half of IT spend at some firms. CIOs saw cloud bills rise 19% in 2025, with generative AI a central driver.
Source: Deloitte, "Navigate the Economics of AI," January 2026
The root cause is uniform model selection.
The spending explosion has a straightforward explanation. Every call in every agentic coding session hits a frontier model. File reads, status checks, linting, formatting, classification, and simple code generation all go to Opus or GPT-5.5. The model does not distinguish between a request that requires deep reasoning and one that could be handled by a model costing 5 to 10x less.
Research on agent token use backs this up. A study led by University of Michigan and Stanford researchers (arXiv 2604.22750) analyzed eight frontier models on SWE-bench Verified. Runs on the same task differed by up to 30x in total tokens, and higher token use did not translate into higher accuracy. Many expensive, failed runs showed repeated file views and re-edits that inflated context without proportional progress. When every call goes to a frontier model, that waste is billed at frontier rates.
The pattern repeats at every scale. Individual developers hit subscription limits and switch to API pricing, where costs balloon further. Teams adopt multiple tools (Claude Code, Cursor, Copilot) and pay for overlapping coverage. Organizations set annual budgets based on 2025 usage patterns and blow through them in Q1 when agentic adoption takes off.
The data on what routing changes.
The fix is not to use cheaper models for everything. Sonnet 4.6 delivers 99% of Opus 4.6's performance on SWE-bench (79.6% vs 80.8%), but there are tasks where that 1.2 percentage points matters. The fix is to stop sending every request to the most expensive model.
Augment Code published a routing guide in 2026 that breaks coding agent workflows into roles: Opus for coordination and architecture decisions, Sonnet for implementation, Haiku for file navigation. Its worked cost model for one hypothetical multi-agent session puts three-tier Claude routing 51% below uniform Opus, $0.98 versus $2.02 per session.
Source: Augment Code, "Best AI Model for Coding Agents in 2026: A Routing Guide," April 2026
AI.cc, an API aggregator, reported a similar pattern across its own enterprise accounts. Between Q1 2025 and Q1 2026, their cost per million tokens fell from $18.40 to $6.07. AI.cc attributes roughly half of that drop (about 34 of 67 percentage points) to multi-model routing, and the rest to open-source pricing pressure and its own volume pricing. The average number of models per account grew from 2.1 to 4.7 in that period, and the tiered routing pattern now covers 64% of its enterprise accounts by token volume.
AI.cc says accounts that fully implemented tiered routing reached a median blended cost of $2.31 per million tokens in 2026, compared with $18.40 for equivalent workloads sent entirely to frontier models twelve months earlier. That is an 87.4% reduction, though it compares two different years and comes from one vendor's customer base.
What the Uber math looks like with routing.
This is an illustration, not a measurement. Take Uber's reported numbers: roughly 5,000 engineers, $150 to $250 per developer per month on average, $500 to $2,000 for heavy users. At the $200 midpoint, that is 5,000 x $200 = $1 million per month.
Now assume 60% of agentic coding calls are low complexity (file reads, simple edits, formatting, classification) and route to Haiku at $1/$5 per million tokens instead of Opus at $5/$25. If those calls use about as many tokens as the rest, the blended cost drops by 0.6 x 80% = 48%, so call it 40 to 50%. That turns $1 million per month into $500,000 to $600,000. The annual difference is $4.8 to $6 million.
That is an assumption-driven estimate, not a measured result. The real number depends on how much of a team's traffic a cheaper model can handle without losing quality, and only outcome-labelled traffic can show that.
The per-developer math is equally direct. A heavy user spending $1,500 per month on uniform Opus usage drops to $750 to $900 with routing. A team of seven spending $3,000 per developer per month drops to $1,500 to $1,800. Even then, a budget that ran out in four months would stretch to roughly seven or eight months (4 / 0.6 is about 6.7, and 4 / 0.5 is 8), so routing is one lever, not the whole fix.
Three things engineering leaders should do now.
1. Audit where tokens actually go. Most teams do not know the complexity breakdown of their agentic workloads. Instrument a week of production traffic. Classify each call by complexity. Many calls will likely turn out not to need a frontier model, but measure your own mix. GitHub instrumented its own agentic workflows and cut effective token use by 19% to 62% in the five workflows it optimized.
2. Route per call, not per session. The model that plans architecture changes is not the model that should read file listings. A trained classifier that evaluates each API call independently and routes to the cheapest capable model captures the gap between frontier and commodity pricing on the calls that do not need a frontier model. Static rules based on prompt length or keyword matching drift as workloads change. A classifier trained on observed outcomes adapts.
3. Set per-developer budgets with routing in the loop. Heavy users can run $500 to $2,000 per month in token costs, so subscription sticker prices are a poor planning basis. DX's guidance runs from $500 to $3,000 or more per developer per year, and a heavy user at $2,000 a month passes the top of that range in two months. Budget at the per-developer level, and include measured routing savings in the forecast rather than assumed ones.
Where Nadir fits.
Nadir can begin beside the gateway in shadow mode. Its Decision API targets about 100 ms, calls no model provider, and returns model, effort, cache, context, and policy evidence. Attach actual usage and coding outcomes before enforcing a cheaper tier; the resulting savings are workload-specific, not assumed from a generic agent mix.
Complete non-streaming responses report the served model and actual cost; a configured, priced benchmark adds the avoided-spend comparison. The dashboard aggregates savings by day, week, and month. No instrumentation needed beyond the two-line change.
For teams whose AI budgets are running ahead of plan, routing is the most direct lever that does not require changing tools, rewriting prompts, or downgrading model quality on the requests that actually need it.
Sources: [The Information, "Uber CTO Shows How Claude Code Can Blow Up AI Budgets"](https://www.theinformation.com/newsletters/applied-ai/uber-cto-shows-claude-code-can-blow-ai-budgets) (April 2026). [Forbes, "Uber Burns Its 2026 AI Budget In Four Months On Claude Code"](https://www.forbes.com/sites/janakirammsv/2026/05/17/uber-burns-its-2026-ai-budget-in-four-months-on-claude-code/) (May 17, 2026). [The Pragmatic Engineer, "How Uber uses AI for development: inside look"](https://newsletter.pragmaticengineer.com/p/how-uber-uses-ai-for-development) (March 2026). [Startup Fortune, "Uber Burned Its Entire 2026 AI Budget in Four Months"](https://startupfortune.com/uber-burned-its-entire-2026-ai-budget-in-four-months-and-claude-code-is-why-finance-teams-should-be-worried/) (May 2026). [The Pragmatic Engineer, "The Pulse: token spend breaks budgets – what next?"](https://blog.pragmaticengineer.com/the-pulse-token-spend-breaks-budgets-what-next/) (April 30, 2026). [DX, "How are engineering leaders approaching 2026 AI tooling budgets?"](https://getdx.com/blog/how-are-engineering-leaders-approaching-2026-ai-tooling-budget/) (October 2025). [Deloitte, "Navigate the Economics of AI"](https://www.deloitte.com/us/en/services/consulting/articles/how-to-navigate-economics-of-ai.html) (January 2026). [Augment Code, "Best AI Model for Coding Agents in 2026: A Routing Guide"](https://www.augmentcode.com/guides/ai-model-routing-guide) (April 2026). [AI.cc, "AICC Report: Enterprise Token Costs Drop 67% Year-Over-Year"](https://www.einpresswire.com/article/911544568/aicc-report-enterprise-token-costs-drop-67-year-over-year-as-multi-model-ai-adoption-hits-record-high) (May 2026). [Bai et al., "How Do AI Agents Spend Your Money?" arXiv:2604.22750](https://arxiv.org/abs/2604.22750) (April 2026). [GitHub, "Improving token efficiency in GitHub Agentic Workflows"](https://github.blog/ai-and-ml/github-copilot/improving-token-efficiency-in-github-agentic-workflows/) (May 2026). Anthropic, OpenAI model pricing as of May 2026.