Per-seat budgets, per-token bills

What changed, with receipts

GitHub announced on 27 April 2026 that all Copilot plans move to usage-based billing on 1 June 2026: premium request units were replaced by GitHub AI Credits consumed against token usage at published API rates. Base plan prices did not move — Copilot Business stays $19/user/month — but each seat now carries a credit allowance ($19 for Business, $39 for Enterprise), credits pool across the organization, unused credits expire monthly, and there is no automatic fallback to a cheaper model when credits run out.

MarketScale reported on 9 July 2026 that Microsoft made Copilot Cowork generally available at $0.01 per credit and that Anthropic moved Claude Fable 5 to usage-based billing after 7 July 2026 while adding enterprise spend alerts — two large AI platforms switching from predictable per-seat costs to variable charges within days of each other. The same article relays Gartner's projection that AI agents will pull $234 billion away from traditional SaaS vendors by 2030.

The practical failure case is well documented in reporting compiled by VarOps on 31 August 2026: Uber capped employee spending on agentic coding tools at $1,500 per person per month, per tool, after its 2026 AI budget was consumed in roughly four months (originally reported by Bloomberg and The Information). VarOps' core observation is the one to keep: vendors re-metered consumption in tokens while buyers kept budgeting in seats.

How big the gap can get

BetterCloud's June 2026 analysis, citing Gartner, forecasts AI-powered application spending near $2.52 trillion in 2026 and warns that scaling pilots to production has produced 500–1,000% cost underestimation in some cases. It also reports that AI add-ons can add 30–110% to base costs and that 68% of vendors kept AI features in premium tiers in 2025.

A Redress Compliance pricing report (a consultancy with an undisclosed methodology — read it as field observation, not measurement) describes an "attach trap": initial AI add-on plans covering 40–70% of seats while measured weekly active use lands at 10–25% of those seats, and consumption lines exceeding budget by 20–60% in year one without caps or alerts.

Bain & Company research, summarized in the VarOps piece, found that across 30+ SaaS vendors adding generative AI, roughly 35% raised per-seat pricing and bundled AI into tiers while about 65% went hybrid (seats plus an AI meter) — and none had fully shifted to usage- or outcome-based pricing at the time of that analysis.

A buyer's checklist for the new meter

None of this argues against AI spend. It argues for the same discipline cloud adoption eventually forced: measure what the meter measures, cap what can run away, and read the invoice as the instrument it is.

• Audit each AI-capable contract before renewal: which are already consumption-based, which are transitioning, and which have protection against mid-term repricing (per MarketScale's guidance).
• Demand usage telemetry by team and workflow before you accept a pool. Pooled credits can hide concentration.
• Set hard caps and spend alerts. GitHub's admin budget controls and Anthropic's spend alerts show the features exist; open-ended usage is a configuration choice, not a requirement.
• Negotiate a usage true-down right on per-seat AI add-ons so attached seats can shrink to measured use (Redress Compliance's central clause recommendation).
• Re-denominate the forecast itself: estimate spend in the unit the vendor meters (tokens, credits, outcomes), then validate against one month of real consumption.