Uber used up its annual AI budget in 120 days. How to pay less for tokens?

Bills for language models are rising faster than company plans, and Uber’s example shows that the main problem is how AI is used, not the price itself. We compare ways to lower costs and check which ones are within reach for a small business.

For two years, companies have been testing what artificial intelligence can do. Now CFOs are starting to calculate how much it costs and what comes of it. The highest-profile example is Uber, but the mechanism that kicked in there threatens any organization billed for AI by usage. We looked into where these expenses come from and how they can be reduced.

How Uber used up its budget in four months

In December 2025, Uber rolled out AI-assisted coding tools for about 5,000 developers: Anthropic’s Claude Code and Cursor. According to “Fortune,” in the first quarter of 2026 the share of engineers using Claude Code rose from 32 to 84 percent. Management accelerated the rollout with internal rankings in which teams were evaluated for how intensively they used AI. So the rankings rewarded the number of prompts, not completed features.

The monthly cost per developer reached as much as $2,000. Sources differ on the starting point: ITwiz writes about a base of $150, while “Fortune” mentions a range starting at $500. According to ITwiz, the company’s CTO Praveen Neppalli Naga used tokens worth $1,200 during a two-hour presentation. In April, the annual AI budget was exhausted. COO Andrew Macdonald admitted on the Rapid Response podcast that he could not connect the growing usage with a 25 percent increase in the number of useful features for riders and drivers. “That link is not there yet,” he said, meaning that connection still is not visible. In June, the company introduced a limit of $1,500 per month per employee and tool, and exceeding it requires a manager’s approval.

Why the token meter blows up budgets

A token is the unit into which a model splits text: a whole short word, part of a word, or a punctuation mark. You pay separately for input tokens (the prompt, a pasted document, conversation history) and output tokens (the model’s response), with the latter usually being more expensive. According to ITwiz, Polish text often uses more tokens than the same content in English.

SaaS licenses (subscription software) got finance departments used to fixed, predictable fees. Token-based billing works like an electricity meter, and AI agents, meaning systems that query the model in a loop on their own and call other tools, spin that meter much faster. Gartner estimates that an agent uses 5 to 30 times more tokens per task than a regular chatbot. Anthropic, based on its own measurements, gives a lower figure: about four times more for a single agent and about 15 times more for a multi-agent system.

Analysts had warned about this earlier. As early as June 2025, Gartner predicted that more than 40 percent of agentic AI projects would be discontinued by the end of 2027, among other reasons because of rising costs and unclear business value. The often-cited 2025 MIT NANDA report says that about 95 percent of organizations did not achieve measurable returns from generative AI. However, the report is preliminary, has not undergone peer review, and is criticized for its small sample size, so this figure should be treated with caution.

First way: a cheaper model where it is enough

On September 29, OpenAI unveiled GPT-6.1 Sol. It costs $2 per million input tokens and $10 per million output tokens, which is five times less than GPT-6 Astra ($10 and $50). Tokens retrieved from cache (previously stored, repeatable parts of prompts) cost $0.10 per million. For example, a team using 100 million input tokens and 20 million output tokens per month would pay $2,000 for Astra and $400 for Sol.

OpenAI says that at a low level of compute effort, the share of answers with factual errors fell from 11.4 percent (GPT-6 Sol) to 7.7 percent and that the difference versus GPT-6 Astra does not exceed 1.9 percentage points. However, these are results from OpenAI’s internal tests. There is still no independent evaluation of the model in company deployments, so before changing the default model, it is worth testing it on your own tasks. Sol is available in Plus, Pro, Business, Enterprise, and Edu plans within ChatGPT Work and Codex.

Uber later took that route. By August, it had quadrupled the number of employees using the most powerful models while at the same time lowering the token cost. Better use of cache, changing the default model, testing new models for efficiency, and showing employees their own usage and hourly cost all helped.

Second way: code and a local model before an expensive API

SEO consultant Chris Green described in Search Engine Journal an experiment with Gemini Nano, a small model built into the Chrome browser and running on the user’s computer. Nano was good at turning collected data into a clear description, but it failed at the more difficult final evaluation, where several signals had to be combined. There, a large model available through an API (an application programming interface through which an app uses a model on the provider’s servers and pays by usage) performed better.

The experiment led to a three-layer split: precise operations are handled by regular code, light interpretation by a small local model, and the large model is turned on only when judgment is needed. Green also notes that the limitations of the small model forced him to improve the code, and that improved the results of large models as well. The same principle works beyond SEO: forms and rigid rules are handled more cheaply by a script or classic automation, while an agent makes sense where there is language, uncertainty, and exceptions.

Third way: route queries to the right model

Routing means that the system decides on its own which model will handle a given query. The peer-reviewed RouteLLM study (ICLR 2025 conference) showed a cost reduction of more than 85 percent while preserving 95 percent of the flagship model’s quality. On top of that comes cache, which according to Anthropic reduces the cost of long, repeatable prompts by as much as 90 percent, and batch processing in the OpenAI API, which is half the price.

Arkadiusz Ilgowski, CEO of NATEK, told ITwiz that companies are increasingly running open-weight models, meaning models with publicly available weights, on their own or controlled servers:

What is available to corporations, and what to a small business

A large organization can deploy its own open models, build routing, and assign limits to groups of employees. Paweł Górka, CIO at Bruk-Bet, suggests separate packages of tokens and tools for developers, office workers, and financial analysts, with a clearly defined goal and training. However, own infrastructure means security, maintenance, and update costs. Ilgowski points out that commercial models remain the better choice where the scale is too small to justify building your own environment.

A small company does not need servers to pay less. A cheaper model as the default is enough, and the flagship one only for difficult tasks. On top of that, cache for repeatable templates, a script instead of a model for simple data processing, and limits set in the provider’s dashboard.

In our view, the most important lesson from Uber is not about the model’s price, but about what the company measures. Teams rewarded for usage use more. So it is worth calculating the cost of one completed thing: a closed ticket, a deployed feature, a handled case. Stanisław Korwin-Kossakowski, CEO of GoNextStage, also advises doing the full calculation only after prompts have been organized, because before that it shows the cost of chaos, not the cost of AI.