On Tuesday June 2, 2026, Bloomberg reporter Natalie Lung published a story headlined Uber Caps Usage of AI Tools Like Claude Code to Cut Costs. The piece, picked up the next morning by Simon Willison, PYMNTS, Outlook Business, and ZeroHedge among others, reports that Uber has set a $1,500 monthly cap on token spending per agentic AI coding tool — Claude Code and Cursor are the two named in the story — after the company exhausted its full 2026 budget for those tools in the first four months of the year.

Three weeks earlier, on May 15, TechRadar's Craig Hale reported that Microsoft's Experiences + Devices division — Windows 11, Microsoft 365, Outlook, Teams, Surface — had told its engineers to be off Claude Code and on GitHub Copilot CLI by June 30, the close of Microsoft's fiscal year. The change was confirmed by developer-tech and the wider trade press over the following ten days. Anthropic's confidential S-1 went in on June 1.

Two of the most cited enterprise reference accounts for Claude Code — one a Magnificent-Seven member, one a Fortune-500 buyer whose CTO has been a public reference for the tool since launch — independently put hard limits on the product in the same month. That is the news.

What $1,500 buys you

The math, courtesy of Willison, is the part worth pausing on. Uber's cap is per tool, so an engineer running both Claude Code and Cursor can spend $3,000 a month on tokens. Annualised, that is $36,000. Levels.fyi puts the median Uber software-engineer total compensation at roughly $330,000. The tool budget, fully spent, comes to ~11% of the salary line. That is not the rounding-error number every AI-investor deck has been quoting since 2024.

The Uber spokesperson statement filed under the cap is the kind of sentence a corporate-comms team writes when the actual policy is we ran out of money:

"We think this is all a pretty straightforward way to responsibly encourage agentic AI adoption and experimentation at scale across the company."

Read past it and you get the underlying numbers. CTO Praveen Neppalli Naga told AI Magazine in April that Uber was "back to the drawing board" on AI tooling spend; he had previously posted that 95% of Uber's roughly 5,000 engineers were using AI tools every month. Adoption ran from 32% in December 2025 to 84% in March 2026. Per-engineer monthly bills, before the cap, ranged from $500 to $2,000. That is the curve the 2025-vintage budget did not survive.

Dara Khosrowshahi's earnings-call line — "around 10% of the company's code is now written and submitted by AI agents" — is the same fact from the other end of the telescope. The 10% is the product. The $1,500 cap is the cost line. Both are now disclosed.

What June 30 ends

The Microsoft side of the same trade has a different shape but lands in the same place. Microsoft did not put a dollar cap on Claude Code. It cancelled the licences.

The internal change set a six-week sunset for Claude Code sessions inside the Experiences + Devices division, with Copilot CLI becoming the "default and preferred coding assistant" for all new projects, in the language quoted across the trade press. Rajesh Jha, Microsoft EVP for Experiences + Devices, gave the official rationale:

"Claude Code was an important part of that learning. At the same time, Copilot CLI has given us something especially important: a product we can help shape directly with GitHub for Microsoft's repos, workflows, security expectations, and engineering needs."

That is a sentence Microsoft has had years of practice writing. Translate it: we cannot run a fiscal year on a vendor whose CLI we do not control.

The gentler half of the story, also confirmed by developer-tech, is that Claude models remain selectable inside Copilot CLI. So the Microsoft–Anthropic revenue pipe is not severed; it has been routed through GitHub. The pricing is now Microsoft's to set. The retention is Microsoft's to engineer. The CLI is Microsoft's to ship. The model is Anthropic's to lose if the next tier of Mythos or Opus does not stay ahead.

Read against the MAI-Code-1-Flash launch on June 2 — Microsoft's own coding model, rolled out to every Copilot tier the same day — the May 15 reporting reads less like a budget memo and more like the first day of a substitution plan. June 30 ends Claude Code as a default. The default that replaces it is the one Microsoft owns the unit economics on.

Why both stories landed in the same week as the S-1

The chronology is the part Anthropic's underwriters will be paid to manage.

• Mid-May — Microsoft's internal Claude Code sunset surfaces in TechRadar.
• May 28 — Anthropic closes the Series H at $965B post-money, ships Opus 4.8.
• June 1 — Anthropic submits the confidential draft S-1.
• June 2 — Bloomberg breaks the Uber $1,500 cap; Microsoft ships MAI-Thinking-1 and MAI-Code-1-Flash.
• June 3 — Anthropic publishes the Claude Partner Network tier framework.

Each of those events is internally consistent. Put them on the same calendar and the story is sharper. Anthropic walked into the S-1 quiet period with two of its most cited reference customers publicly imposing cost ceilings on its agentic coding product, and walked out forty-eight hours later with a partner-tier framework that hands enterprise integration to Accenture, Cognizant, Deloitte, KPMG, Infosys, and PwC — a combined ~1.4 million billable people sitting between Anthropic and the customers it cannot, by law, do retail roadshow on right now.

The customer-concentration disclosure in the S-1 will eventually surface what fraction of Claude Code revenue runs through Microsoft- and Uber-scale accounts. The pricing-power disclosure will surface how usage-based billing — the model Anthropic moved Claude Code onto earlier in 2026 — survives a customer base that has now learned to put a meter on it. Neither disclosure is in the draft as currently filed. Both are in the staff comment letters that come next.

What this means

Three takeaways.

1. Usage-based billing finds its ceiling when customers print their own. Uber's $1,500 cap and Microsoft's June 30 sunset are different mechanisms doing the same job: they convert a metered cost into a budget line. That is the move every CFO eventually makes on every line item that started life as "$500 to $2,000 per engineer per month." Anthropic's revenue model assumed the meter ran upward as agents got better. The agents did get better — Khosrowshahi's 10% AI-written-code number is the proof — and the meter ran upward, and a Fortune-500 buyer and a Magnificent-Seven member capped it in the same month. The ceiling is now in the data, not just the discussion.

2. Microsoft's exit is the more strategically dangerous of the two. Uber is a customer; the cap is a budget signal, not a vendor swap. Microsoft is a customer and a competitor, and the May memo is the operational predicate for the June 2 MAI-Code-1-Flash launch. The substitution will not be visible in Microsoft's quarterly Copilot disclosures until it has already happened, because Claude is still selectable through the CLI Microsoft now controls. Anthropic's S-1 risk-factor language will list this scenario. Investors will price it the way they priced Snowflake-vs-Databricks: not as zero, but discounted from the gross-revenue line that gets printed at the top of the deck.

3. The Partner Network is the channel Anthropic just built to route around its largest accounts. Three months ago the enterprise sales motion ran through cloud partners and direct relationships with names like Microsoft and Uber. As of Wednesday it runs through six global systems integrators with a combined million-plus billable workforce, certified against three published tiers. That is what a company does when the top of its customer pyramid has either learned to negotiate hard or started building substitutes. The S-1 will not say "we are diversifying go-to-market because the top of our funnel has started capping the meter." It will say something about "deepening partner-led implementation." Same sentence, two readings.

The S-1 is confidential. The caps and the deadline are not. The first chart in the public S-1 — the one with the customer-revenue mix and the cohort-retention curve — is the chart everyone in the AI-IPO book is now waiting for. Anthropic chose its filing window. It did not choose this week.