Build1 publisher2 min readPublished
Qodo caps engineer tokens at $10,000 a month, a level most of its developers never reach
CEO Itamar Friedman says the ceiling exists to force someone to answer which automation is worth the money. The bill growing faster is the infrastructure Qodo runs for customers, up roughly 5x year over year.
The Engineer · Build desk
What happened
- Qodo CEO Itamar Friedman told The New Stack that his engineers can draw $10,000 worth of tokens a month, and that most of the company's developers never reach that ceiling.
- Friedman says the cap was not enacted to restrict usage but to drive visibility and efficiency so the startup can scale without runaway costs.
- Claude Code is still the most used tool among Qodo's own engineers, though staff are shifting quickly towards OpenAI's Codex.
Compiled by The EngineerSomething wrong?How this is made
Why it matters
- decision Because the ceiling is per seat and visible during the month, an engineer who wants an agent running unattended has to make the case for it before the invoice arrives.
- constraint One ratio of all benefits over all costs cannot attribute return to a single assistant, so anyone wanting per-tool ROI has to pair usage data with a preference signal.
- cost The part of Qodo's AI bill that scales is the customer-side infrastructure, and it keeps scaling whatever the per-engineer token cap is set to.
- contradiction Pull request volume measures the writing speed Friedman says was already solved. That leaves the bug and incident line as the only half of his pair that tests software velocity.
Qodo's ceiling is per seat and it sits above where most seats land, so it does not ration consumption; it puts a number on the one workload that will cross it. Qodo's own explanation for why its customer-side bill climbs names that workload: increased user adoption, plus agents taking on "more, and longer tasks" as they mature [6].
Qodo has 130 staff spread across offices in several countries [11]. At $10,000 each [1], that is $1.3 million a month and $15.6 million over a year [18]. That is a little over a fifth of the $70 million Series B the company raised earlier this year [12][19]. Friedman called the cap "generous" [16]. Only some of those 130 are engineers, so the authorised figure sits well above any plausible draw.
What the ceiling is for, per Friedman, is making somebody answer one question: which path of automation or usage would be the best use of the company's money [4]. The ROI test he recommends is a ratio, AI benefits over AI costs, where a larger number means a better return [14]. He took the shape from The Phoenix Project, the 2013 DevOps novel, and puts features and infrastructure in the numerator [15]. The per-tool comparison runs on separate instruments: Friedman polls the 130-person staff on the tools they prefer, then checks those answers against what the usage data shows [11].
Six doublings in a year is 64x [20]. Qodo reports pull request volume roughly doubling every couple of months [8], and infrastructure spend growing roughly 5x year over year [5]. That works out to cost per pull request falling by about 13x [21]. It would line up with the company's claim that it is driving down the cost of a reviewed pull request through routing and inference efficiency [7]. For the 13x to mean anything, the two rates have to cover the same population over the same window. The definition of a pull request also has to hold steady while agents write more of them. The New Stack's account reports both as rates with no absolute counts, and leaves open whether the pull request figure covers Qodo's own repositories or customer usage [22].
"We solved the speed of writing code," Friedman said, "we didn't solve the velocity of creating software" [9]. Qodo runs its own pull requests through Qodo, and Friedman said the product powers the company's entire software development life cycle [17].
What to watch
- Whether Qodo publishes the absolute pull request and incident counts that would let the 13x per-PR figure be checked.
- Whether Codex passes Claude Code in Qodo's internal usage data as well as in the staff preference poll.
- Whether the routing and inference efficiency work bends the 5x infrastructure growth rate as agents take on longer tasks.