September 20, 2026
Unlimited AI budgets are over: expensive models are kept only for planning
A company with an unlimited AI budget has introduced a daily cap on its smartest models: Fable and Astra are reserved for planning.
Gergely Orosz
@gergelyorosz
I spoke with a company where AI budgets were unlimited a year ago, and whose CEO is a technologist and a major AI enthusiast. “Now we have a daily budget cap for SOTA models. Fable and Astra are only for planning; cheaper models are enough for everything else.” Let me stress: this is a company steeped in AI, and it never used to look at AI budgets at all.
· 234.3K views
A year ago, no one at this company counted the AI budget. Now an expensive model does the thinking, while a cheap one writes the code.
The cap was imposed not by skeptics, but by a team that believes in AI and has been paying for it since last year. The drive to cut costs has reached even the most committed believers.
They used to pay as much as they wanted. Uber encouraged employees to use AI as much as possible and ranked spending on internal leaderboards; individual engineers spent from $500 to $2,000 a month. The company burned through its annual AI budget in 4 months and, by April 2026, set a ceiling of $1,500 per person per month for every agentic tool (TechCrunch, 02.06.2026).
Since July 2026, Microsoft has given divisions target token budgets and switched its default internal model to a cheaper one. EVP Jay Parikh wrote to employees that the company is not optimizing for token spend (The Next Web, 04.08.2026).
The setup is not new. As early as April 2026, developers described the same approach: start in planning mode on Opus, then hand execution to Sonnet; around 30% of respondents hit the limits of $100–200 max plans per engineer per month (The Pragmatic Engineer newsletter, 14.04.2026). The math is simple: Sonnet 5 is five times cheaper than Fable 5.1 for output, Haiku 4.5 is ten times cheaper; OpenAI has the same gap, with Sol 60% cheaper than Astra.
The setup is built in. In Claude Code, it is configured with a single alias: the `/model opusplan` command keeps Opus in planning mode and automatically switches to Sonnet when the agent starts writing code. Cheaper models for everything else are set with the `model: haiku` line in a subagent's frontmatter and a list of allowed models in settings.json (Claude Code docs, 20.09.2026).
The cap itself is set at three levels. A session is limited with the `--max-budget-usd` flag. Team and individual spending is cut with spend limits in the admin panel on Team and Enterprise plans, while an API organization is limited through workspace limits in the console. The same docs provide a benchmark: around $13 per developer per active day, $150–250 a month, with 90% of users spending less than $30 per active day.
Next, the “model budget” line moves from an experiment into a standard team budget: it has been there at Uber and Microsoft since summer.
