AI Got Cheaper. Microsoft and Uber's Bills Went Up Anyway.

1 / 9

Cheaper Tokens, Bigger Bills — How Both Are True

AI output token prices have fallen more than 80% since 2023 across leading models. Yet in 2026, two of the world's most sophisticated tech companies — Microsoft and Uber — both hit real budget walls using AI coding tools. The explanation isn't that AI got expensive. It's that flat-rate pricing was never built for how agentic tools actually get used.

OECD AI Markets Report 2025 (via IntuitionLabs analysis)

Swipe or use ← → to navigate

AI Got Cheaper. Microsoft and Uber's Bills Went Up Anyway.

Token prices have fallen sharply since 2023. Yet Microsoft cancelled its own engineers' access to Claude Code over cost, and Uber burned a year's AI budget in four months. Khan Academy's founder says the same rising bill is the best money his nonprofit spends. Here's what actually happened on both sides, sourced to the reporting and the people who said it.

16 July 2026

AI Got Cheaper. Microsoft and Uber's Bills Went Up Anyway.1 / 9

Cheaper Tokens, Bigger Bills — How Both Are True

OECD AI Markets Report 2025 (via IntuitionLabs analysis)

AI Got Cheaper. Microsoft and Uber's Bills Went Up Anyway.2 / 9

The First Crack: GitHub Paused Its Own Signups

In November 2025, GitHub paused new sign-ups for Copilot Pro and Pro+ — because agentic workloads from existing paying customers were generating costs that exceeded what those customers were actually paying. Flat-rate pricing was losing GitHub money on its own heaviest users. That's the moment the old model broke, months before anyone outside the industry noticed.

The Next Web — "Microsoft's quiet Claude Code retreat and the real cost of enterprise AI"

AI Got Cheaper. Microsoft and Uber's Bills Went Up Anyway.3 / 9

The Industry Response: Pay By The Token, Not The Seat

GitHub moved all Copilot plans to usage-based billing on June 1, 2026 — replacing flat premium-request limits with token-linked "AI Credits" priced per unit of actual compute used. GitHub's own reasoning was direct: Copilot is no longer lightweight autocomplete, and agentic workflows consume far more compute than a flat seat price was ever designed to absorb.

GitHub Copilot — official plans and pricing page

Forbes — "Why Your Engineers' Favorite AI Tools Are Wrecking Your 2026 Budget" (Janakiram MSV)

AI Got Cheaper. Microsoft and Uber's Bills Went Up Anyway.4 / 9

Microsoft Let Its Own Engineers Choose — Then Took It Back

In December 2025, Microsoft opened Claude Code to thousands of its own engineers across the team that builds Windows, Microsoft 365, Outlook, Teams, and Surface. It became the preferred tool in that group. Then the bill arrived — usage ran past the annual AI budget months ahead of schedule. Microsoft is winding down access by June 30, 2026, the last day of its fiscal year, and moving those engineers to its own GitHub Copilot CLI instead.

Forbes — "Microsoft Ends Claude Code Licenses As It Shifts Developers To Copilot" (Jon Markman)

Fortune — "Microsoft reports are exposing AI's real cost problem"

AI Got Cheaper. Microsoft and Uber's Bills Went Up Anyway.5 / 9

Uber Burned a Year's Budget in Four Months

Uber's Chief Technology Officer, Praveen Neppalli Naga, told The Information in April 2026 that Uber had burned through its entire planned 2026 AI coding budget in just four months. The company hadn't stumbled into it passively, either — Uber had run internal leaderboards ranking engineers by their Claude Code activity. The incentive worked exactly as designed. The budget didn't survive it.

Forbes — "Why Your Engineers' Favorite AI Tools Are Wrecking Your 2026 Budget" (Janakiram MSV, citing The Information)

AI Got Cheaper. Microsoft and Uber's Bills Went Up Anyway.6 / 9

This Isn't Just Microsoft and Uber

A 2025 study from cost-governance firm Mavvrik and Benchmarkit, covering 372 enterprises, found only 15% of companies forecast their AI costs within 10% of actual spend. A majority miss by 11-25%, and nearly one in four miss by more than 50%. Worth noting: Mavvrik sells AI cost-governance tooling, so this is relevant evidence, not neutral academic data — but the pattern lines up with exactly what happened at Microsoft and Uber.

Forbes — "Why Your Engineers' Favorite AI Tools Are Wrecking Your 2026 Budget" (Janakiram MSV, citing Mavvrik/Benchmarkit study)

AI Got Cheaper. Microsoft and Uber's Bills Went Up Anyway.7 / 9

This Curve Is Still Accelerating

Goldman Sachs has forecast that agentic AI could drive a roughly 24-fold increase in global token consumption by 2030, reaching an estimated 120 quadrillion tokens per month, as both consumers and enterprises adopt AI agents more deeply. If usage keeps compounding at anything close to that rate, the Microsoft and Uber stories aren't cautionary tales from an awkward transition year — they're the first of many.

Fortune — "Microsoft reports are exposing AI's real cost problem" (citing Goldman Sachs forecast)

AI Got Cheaper. Microsoft and Uber's Bills Went Up Anyway.8 / 9

The Other Side of the Same Coin: When the Spend Is Worth It

Khan Academy founder Sal Khan said on the Silicon Valley Girl podcast that the nonprofit's Anthropic bill runs at roughly $1.2 million a year, for a team of around 200, mostly engineers — and it's growing fast. He relayed that Khan Academy's CTO reported one engineer spent $3,000 in a single day on compute, which let them complete in a few hours what would normally have taken three to four months. Khan's own framing: producing what now costs $1-3 million a year in tokens would have cost $20-30 million to build the old way. Same rising bill everyone else in this deck is fighting — but here it's treated as the best money the organization spends, not a runaway cost to cap.

Sal Khan, Silicon Valley Girl podcast with Marina Mogilko (as stated in the episode)

AI Got Cheaper. Microsoft and Uber's Bills Went Up Anyway.9 / 9

What To Actually Do About It

This is Dimension 2 from the model-selection framework playing out in real time: token price is one line in a bigger bill. Before you scale an agentic coding tool org-wide:

Ask your vendor directly whether pricing is still flat-rate, and when that changes
Track actual token burn per developer for 2-4 weeks before a team-wide rollout
Model your worst-case monthly bill, not your average case — agentic usage is spiky, not steady
Set a hard budget alert threshold, not just an annual figure — Uber's leaderboard shows how fast incentives can outrun a budget
Revisit vendor pricing quarterly — GitHub, Anthropic, and OpenAI all moved in the same six-month window

Synthesized from the reporting cited on the preceding slides