Uber reportedly exhausted its AI budget on coding agents, and input tokens are why

A reported budget overrun at Uber

A post on dev.to from Sentra, a startup building a memory layer for AI agents, reports that Uber's CTO said the company had exhausted its AI budget a few months into 2026, with coding-agent usage cited as the main driver. The claim is second-hand — the author says the remark was reported rather than made to them directly — and no primary source is given, so it should be read as an unverified account rather than a confirmed figure. Within the post, Uber is framed as the first large engineering organisation to hit token-cost limits in public, and that framing is what makes the anecdote useful even without confirmation: the cost pattern it describes is one any team running agent fleets can check against its own invoices.

Where the money actually goes

Most of the essay is arithmetic. Using Anthropic's published per-million-token rates as of August 2026 — the post lists Claude Fable 5 at $10 input and $50 output, Opus 5 at $5 and $25, Sonnet 5 at $2 and $10, and Haiku 4.5 at $1 and $5 — it models a small fleet: five agents, 200 model calls per agent per day, 20,000 input and 1,500 output tokens per call, across 30 days. That totals 600 million input tokens against 45 million output tokens, or $4,125 a month at Opus pricing, with about 73% of the bill coming from input.

The conclusion the author draws is that agent spend is dominated not by what models generate but by the context resent on every call, and that most cost-control effort targets the smaller output side of that ratio.

Four levers, ranked by effect

The post ranks four reduction levers:

• Prompt caching lowers price, not volume. Cache reads are billed at 0.1x the base input rate, with writes at 1.25x for a five-minute window or 2x for one hour, per the post's figures. In the worked example that saves roughly 9%. The author notes a caveat: caching is indifferent to whether cached content is still correct, so it is a price lever, never a correctness lever.
• Model routing is presented as the biggest single line item. The same fleet on Haiku 4.5 instead of Opus 5 drops from $4,125 to $825 a month. The author says teams resist this longest and should route by task difficulty rather than habit.
• Batch APIs cut the price of anything asynchronous in half. According to the post, both major providers discount batch input and output by 50%, making it a configuration change worth exactly half the bill on traffic that tolerates latency.
• Sending less context is called the only lever that compounds, because the first three lower the price of tokens while this one reduces how many are needed. The post contrasts a retrieval bundle of four documents totalling roughly 7,000 tokens, two of them contradictory, with about 200 tokens of resolved facts carrying validity and provenance.

A benchmark worth a caveat

To support the fourth lever, the author cites Sentra's own evaluation on Terminal-Bench 2.1: agents given a task-scoped memory layer used 41.2% fewer tokens at 72.6% lower model cost, while accuracy rose from 83.37% to 88.31% mean reward across 445 trials. The post itself flags this as its own measurement and tells readers to examine the methodology rather than trust the number. It is also worth noting that Sentra sells exactly the kind of memory product this lever implies, so the benchmark should be read as vendor-supplied evidence for the least verifiable claim in the piece. The author's explanation for cost falling faster than token use — fewer retries and shorter runs compounding with smaller payloads — is plausible but self-reported.

A practical checklist

The post's ordered recommendations for teams: split the bill into four buckets (system prompt, retrieved context, conversation history, output) to find which dominates; delete dead weight from the system prompt, which is billed on every request; enable caching for stable prefixes and move variable tokens to the end of the prompt; cap conversation history, since resending a full transcript each turn grows cost quadratically over a session; route easy work to smaller models; then check whether agents are re-deriving the same context every session, which no amount of caching fixes.

Why it matters

If the Uber account is even roughly accurate, it is an early public preview of a budget shift engineering organisations are only starting to model: spend moving toward metered token consumption driven by coding agents, with input context as the dominant cost. The post coins the term 'tokenmaxxing' for the strategy of maximising token use and trusting value follows, and argues it is not foolish — it is a rational first response to genuinely useful tools — but that spend should target relevance rather than volume. The cynical reading, that a memory vendor is dressing a product pitch as a cost crisis, does not make the underlying arithmetic wrong: the input-heavy cost ratio and the routing and batching levers are checkable against published pricing by anyone with an invoice.