Qodo, a startup that uses artificial intelligence to ensure AI-generated code meets quality and governance standards, has placed a $10,000 monthly limit on the tokens its engineers can access, according to a profile published by The New Stack. CEO Itamar Friedman characterized the ceiling as "generous," noting that most developers at the company never hit it. The cap wasn't introduced to "restrict usage," Friedman said, but rather to create "visibility and efficiency" so the startup can "scale without runaway costs." The company raised a $70 million Series B earlier this year.
Qodo's AI infrastructure spending—the expense of operating the product for customers, separate from internal engineer token use—is climbing at "roughly 5x year over year," the company told The New Stack, driven by "increased user adoption" and agents handling more and longer tasks as they develop. At the same time, Qodo reports it's reducing the cost of reviewed pull requests through routing and inference efficiency. The startup, which employs 130 people across multiple countries, is seeing "roughly double" the number of pull requests "every couple of months" alongside "a decreasing amount of bugs and incidents," Friedman reported. Internally, Claude Code remains the leading coding assistant, though OpenAI's Codex has been gaining ground, with staff "shifting quickly towards Codex." Friedman tracks tool preferences both through polling his team and by analyzing usage data.
Friedman recommends an equation-based method for measuring AI return on investment, comparing all benefits against all costs to produce a ratio where larger results signal greater returns. He discovered the structure of the calculation in *The Phoenix Project*, a 2013 DevOps novel that contrasts categories of software development work and divides them into good and bad groups. According to Friedman, "what you can't measure, you can't improve," and he contends that an imperfect measurement beats having none at all. He advises startups to select no more than six or eight terms for their own calculations and not to "think about it too much"—just "try to put any number [in the AI ROI equation] and start tracking." The company went all in on automating email, then retreated, "mov[ing] from AI automation to AI enhancement" after finding that AI struggled to match writing tone and intelligently pull tasks from messages.
The difference between accelerating one step of a task and its entire process is the gap between AI hype and AI ROI, the report argues. Companies adopting AI coding tools often discover they generate more code with machines than their human staff can review, simply moving the bottleneck one step down the software development life cycle. "We solved the speed of writing code," Friedman contends, "we didn't solve the velocity of creating software." Without current context, AI can do little more than "filibuster," he noted, and automation falls short today in two areas: when human judgment is required and when context is missing. Qodo's stated method for any task is to "go all in on complete automation," then "take a step back to human judgment." The company doesn't mandate AI use among employees—workers won't be fired simply for not being "AI all the way" or "eating AI for breakfast"—but staff are expected to complete work as efficiently as possible, and those who lag on assigned tasks are expected to turn to automation. Friedman advises founders to predict "what's going to happen two years from now" and solve for it immediately, because whatever seems two years away tends to arrive within twelve months. The future "is coming faster" than you think, he said, admitting that predicting ahead is difficult "but you have to." For decision-makers, the tension between allowing autonomy and enforcing efficiency through automation will only sharpen as tools mature and performance gaps widen. The real discipline lies not in adopting every new capability, but in knowing which costs to accept and which efficiencies to reject.