The GitHub Copilot community FAQ has 904 downvotes and 22 upvotes. That sums up the transition that happened between March and June 2026: every AI coding tool migrated to metered token billing. Flat-rate is dead. And the real numbers are more dramatic than any projection.
The mass migration to metered billing
Between March and June 2026, six major tools migrated from flat per-seat to metered token billing:
- GitHub Copilot (June 1): swapped Premium Request Units for AI Credits. When credits run out, the service stops. Pro: $10/month with 15 credits. Pro+: $39 with 70. Max: $100 with 200.
- Cursor (June 1): restructured Teams with dual usage pools. Standard $40/seat, Premium $120/seat with 5x usage.
- Windsurf (March 19): replaced credits with daily/weekly quotas. Pro $15-20, Max $200. Rebranded as Devin Desktop in June.
- Anthropic (April 15): enterprise migrated from flat per-seat to per-token with mandatory commitment.
- Claude Code (June 15): billing split — interactive sessions stay on subscription, programmatic usage (Agent SDK) goes to a separate credit pool at full API rates.
- OpenAI Codex (April 2): from per-message to token-based.
According to Gartner, flat-rate was a loss leader that built adoption. Adoption locked in, real billing began.
The real numbers
The cost multipliers are staggering. According to reports compiled from Reddit, Business Insider, and Gartner:
- Reddit user: from $29/month to ~$750 (~26x).
- Another Reddit user: from $50 to ~$3,000 (~60x).
- Pro+ subscriber projected at $847/month, according to Business Insider.
- Visual Studio Magazine editor: burned through 82% of 1,500 free credits on Day 1.
- Uber: $500-2,000 per engineer/month (up from ~$150-250).
- According to Gartner, accounts jumping from $20-100 to $2,000-5,000 per developer/month, with extreme cases at $20,000.
The essential caveat: these multipliers come from heavy agentic users. Code completions and Next Edit remain free on Copilot. Casual chat users feel little impact. Those running multi-hour autonomous sessions feel everything. The difference between casual chat and autonomous agentic work is the difference between a coffee and a rent payment.
Jevons Paradox in practice
The mechanism behind the cost explosion is Jevons Paradox. Per-token prices dropped ~80% in 2025-2026. Total AI spend went up. Why? Agentic architecture is a token multiplier: more turns per task, more tokens per turn. A single task can consume 1-3.5 million tokens. Power users spending $1,800+/month.
According to GitHub CPO Mario Rodriguez: a quick chat question and a multi-hour autonomous session can cost the user the same. The current premium requests model is no longer sustainable. When the marginal cost of an autonomous session is 100x the cost of a chat, flat-rate becomes a hole in the balance sheet.
The market response
The Linux Foundation launched the Tokenomics Foundation (June 3, 2026) at FinOps X, expanding the FOCUS spec for token-based spend. According to Gartner, AI coding costs will exceed the average developer salary by 2028. In India, token costs already equal a 4-6 year engineer's salary. 63% of organizations implementing spend controls. Uber imposed a $1,500/month cap per engineer. Microsoft is canceling Claude Code licenses by June 30. According to Goldman Sachs, annual AI infrastructure spend could rise from $765 billion (2026) to $1.6 trillion (2031).
Token discipline will not emerge from developer choice alone, according to Gartner. FinOps for AI coding is the next frontier.
What we see at Tech86
At Tech86, we have been tracking this transition closely. The pattern is always the same: a team adopts copilots, usage grows organically, nobody monitors spend, and the bill surprises at the end of the month. The 26-60x multipliers are not exceptions — they are what happens when developers discover autonomous sessions without budget guardrails.
The solution is not to cut AI — it is to implement FinOps for AI with the same discipline you already use for cloud. Model-tier routing, feature-level budget guardrails, and team-level accountability. Flat-rate is dead, but the productivity it unlocked is real. The challenge now is to pay for what you use — and use what you pay for.