Should Juniors Get Fewer AI Tokens? Apple Caps Some at $200

A principal engineer at a publicly traded company has proposed that junior engineers “have much lower token budgets than senior engineers”, Factory’s Dex Horthy said on the MTS live show, in a clip MTS posted on 17 November. Horthy did not name the engineer and said: “I do not endorse this, but it is a potential path that could be explored”.

Companies already ration AI spend, though not by seniority in any report we have covered. Bloomberg reported on 4 November that Apple cut some engineers’ daily AI budget to $200 from $300 and that Tesla set a $200 weekly limit. Our number: a top-1% user on Databricks’ Unity Gateway spent $699.74 in the week of 2 November, 3.5 times that Tesla limit and 281 times the median user’s $2.49, by our arithmetic.

Business Pill · Usage-Based Pricing

An evergreen Business Engineer explainer on usage-based pricing, paying for work rather than seats. It teaches the general idea only and does not discuss Horthy, MTS or the companies named here.

The key insight: As we read it, the proposal Horthy relayed would add seniority to AI budgets that companies already set by team and tool: Bloomberg reported Apple’s $200 daily cap for some engineers and Tesla’s $200 weekly limit, while a top-1% Databricks user spent $699.74 in the week of 2 November, 3.5 times that Tesla limit, by our arithmetic.

What Horthy Said on MTS

Host Madison Kanna wrote on X on 16 November that she would be talking with Horthy on MTS that day about “the AI no slop manifesto”. In the clip she puts a viewer’s question to him: “if engineers traditionally develop judgment through writing, reading, and debugging code, how will the next generation develop that judgment?”

Horthy answered that “people ask me this all the time and I don’t think I have a good answer”. He then relayed the proposal from a friend he described as a principal engineer at a publicly traded company with “probably 500 engineers”.

In the version he described, “maybe if you’re more junior, you just get the tab complete model until you’ve like leveled up”, and after that “you get to start using one agent and then you get to start using massive swarms of agents”. Horthy added: “I don’t know if I endorse that as a solution”.

AI spend per user per week on Databricks’ Unity Gateway at the median, 90th and 99th percentiles in the
AI spend per user per week on Databricks’ Unity Gateway at the median, 90th and 99th percentiles in the week of 2 November 2026, from the Databricks AI Index, against the $200 weekly limit Bloomberg reported at Tesla. The 3.5x and 281x ratios are our arithmetic.

What Was Posted on X

Posted by the MTS live show at 04:50 UTC on 17 November, this 168-second clip has Factory’s Dex Horthy answering a viewer’s question about how new engineers will build judgment. He relays an unnamed principal engineer’s proposal to give junior engineers lower token budgets and says he does not endorse it. Every quote in the piece comes from our own transcript of the clip, not from the caption.

The Problem Horthy Described

Horthy described the problem through code review. A comment that sends a junior to see how a team already does something, instead of handing over the fix, used to teach the pattern. Today, he said, the person given that comment is “just gonna paste it into their agent”, and “the agent’s gonna do that and they’ve learned nothing”.

He described an approach he has seen from Mitchell Hashimoto: “He literally says, talk to your agent until you understand why this is wrong”. Juniors are hired, Horthy said, not for day-one output but because “they’re gonna be seniors someday”.

Budgets Already Exist, by Team and by Person

Bloomberg reported on 4 November that many groups outside engineering at Apple already face a $100 daily cap and that Apple has considered as little as $25 for some teams (our piece). Between the $200 engineering cap and that $25, the spread is 8 times, by our arithmetic. Employees who need more “will be able to seek exemptions from senior leadership”, Bloomberg wrote; Apple declined to comment.

Bloomberg also reported that Uber capped individual AI coding tools at $1,500 per month, that Meta introduced token budgets and that Microsoft tightened internal AI spending allowances. Goldman Sachs limited Claude Code and GitHub to its engineers, a limit by role (our piece).

Inside companies, spend per person is already uneven. On the Databricks AI Index, the median user spent $2.49 in the week of 2 November, the 90th percentile $95.96 and the 99th percentile $699.74 (our piece). Databricks co-founder Patrick Wendell wrote that “Power user spend appears to be holding mostly flat, likely as a result of budgeting, while the median user spend is quickly growing from a small base”.

Across firms, the gap is wider. On Ramp’s AI Index, the top 1% of firms paying for AI spent a median $10,199.20 per employee per month in October, against $16.32 for the median firm, about 625 times (our piece).

Apple daily AI spending caps as reported by Bloomberg on 4 November 2026: some engineers $300 to $200, some engineering teams $300 to as little as $100 planned, some teams outside engineering $100 to $25 considered
Apple’s daily AI spending caps per employee as Bloomberg reported them on 4 November 2026, citing people with knowledge of the changes: the cut for some engineers, the further cuts planned for some engineering teams and the $25 Apple has considered for some teams outside engineering. Apple declined to comment.

The Model Is a Budget Too

The ladder Horthy described runs on model choice as much as on dollars. On the coding benchmark on the Databricks AI Index page, four entries run in Claude Code scored 89.1 on quality: Claude Opus 5.5 at $1.616 per task, GLM 5.3 Flash at $0.178, GPT 6.1 Terra at $0.12 and GPT 6 Luna at $0.053.

For the same score, Opus 5.5 cost about 30 times as much per task as GPT 6 Luna, by our arithmetic. At those costs, $200 buys about 124 Opus 5.5 tasks or about 3,774 GPT 6 Luna tasks.

Anthropic says about 90% of the Claude Code team’s code is written and reviewed by Claude (our piece). Earlier we covered Ryan Dahl’s view of AI-written code and Menlo Ventures’ AI coding market figures.

The Structural Read

As we read the clip, the problem Horthy describes is learning, not cost: a review comment that used to send a junior to study a pattern can now be pasted into an agent.

As we read the Databricks benchmark, a model tier and a dollar tier are different levers: four entries with the same 89.1 quality score cost from $0.053 to $1.616 per task, about 30 times apart, by our arithmetic.

Dex Horthy, Factory, on MTS

“He literally says, talk to your agent until you understand why this is wrong”

Three Implications

FOR ENGINEERING MANAGERS Bloomberg reported that Apple employees who need higher allowances “will be able to seek exemptions from senior leadership”, an approval path by request rather than by level.

FOR FINANCE TEAMS Wendell wrote that power-user spend appears to be holding mostly flat, “likely as a result of budgeting”, while the median user’s weekly spend rose from $0.40 to $2.49 between the weeks of 20 July and 2 November.

RELATED Our piece on Ramp’s index covers the $10,199-a-worker top 1% of firms.

The Business Engineer Lens

This story maps onto the Business Engineer framework The CFO’s Guide to the Token Economy.

The framework argues that “As companies move from chatbots to thousands of autonomous agentic workflows running 24/7, the line item that used to be a rounding error becomes the budget.”

As we read it, the ladder Horthy described, from a tab complete model to one agent to swarms of agents, is that move inside a single engineering team, with the budget decided person by person. The framework predates the clip and does not discuss it.

What Is Not Established

The proposal is an unnamed engineer’s idea, relayed by Horthy, who does not endorse it; we could not identify the engineer or the company. It is not a policy any company has said it uses. The budgets Bloomberg reported are by team, role or tool, as we covered them.

Databricks measures spend per user on one gateway at list prices, and Ramp measures spend per employee across firms; the two are not the same population. The benchmark costs are Databricks’ figures for its own coding task set. The quotes come from our own transcript of the 168-second clip, part of a longer conversation; we did not contact Horthy, Factory or MTS.

This Week In The AI Supercycle

AI Lost Its Central Banker, the Lab Oligopoly, and the Future of Open Source

This week’s issue: the Nvidia leadership change, read as the AI build-out losing its central banker: who now finances the GPU landlords, backstops the data centres and sponsors open models, and why two labs are becoming the bank.

Read the full issue →

Business Engineer Framework

The CFO’s Guide to the Token Economy

A Business Engineer framework on why cheaper tokens still produce bigger AI bills.

Read the premium analysis →Get 50% off →

The Bottom Line

An unnamed principal engineer’s idea, relayed by Horthy, would tie AI token budgets to seniority. The limits Bloomberg reported so far are by team or tool, and a top-1% Databricks user’s $699.74 week is 3.5 times Tesla’s reported $200 weekly limit, by our arithmetic.

95,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.

A note on sourcing. Dex Horthy’s words come from our own transcript, made in two passes, of the 168-second clip MTS posted on X on 17 November 2026; we quote only what both passes and the audio support, and the proposal is an unnamed principal engineer’s idea that Horthy relayed and said he does not endorse. The company budgets are as Bloomberg reported them on 4 November, which we covered at the time from Bloomberg’s full article.

Spend per user comes from the Databricks AI Index and Patrick Wendell’s post of 16 November, spend per employee from the Ramp AI Index, both read in full. Ratios and task counts are our arithmetic. We did not contact Horthy, Factory, MTS or the companies named.

Sources: MTS clip with Dex Horthy on X, 17 Nov 2026 (our transcript) · Databricks AI Index · Patrick Wendell on X, 16 Nov 2026 (post 4 of 6) · Ramp AI Index · Madison Kanna on X, 16 Nov 2026

Scroll to Top
FourWeekMBA