The all-you-can-eat era ends — AI coding tools learn to meter the margin
Flat-rate subscriptions worked while inference was cheap relative to the fee. Agentic coding broke that math, and the bill is now being handed back to the developers generating it.
GitHub Copilot moved to usage-based billing — a credit a cent, a $100 Max tier, agentic browser tools on by default — while Cursor split its seat pools and added a $120 Premium seat. The billing change is an admission about cost structure: agentic features burn large numbers of tokens, and a fixed monthly price stops covering the heavy users.
Two companies, one pressure
Copilot metering usage and Cursor separating first-party from third-party model consumption are the same reckoning handled two ways. Both are re-architecting pricing around the fact that agentic coding is expensive to serve, and both are targeting the power users whose consumption broke the flat tier. When the two leading tools move to meter in the same month, the flat-rate model is not being tweaked — it is being retired.
Why developers will pay it anyway
Because they are not choosing one tool — they are stacking several. 59% of developers run three or more AI coding tools in parallel, and Claude Code leads satisfaction at 46% "most loved" against Copilot's 9%. A market where developers keep multiple specialised tools and rate them that highly is a market that will absorb metered pricing, because the tools have become genuinely useful rather than merely novel.
The satisfaction gap doubles as a business signal: the same Claude Code developers rate highest is the surface analysts credit for propelling Anthropic's revenue past its rivals. Developer love and enterprise revenue are, for once, the same number — which is why the coding layer is a durable, differentiated market rather than a winner-take-all race.
The same dynamic, one layer up
Zoom out and this is the frontier price war reflected in the tool layer. As per-token cost falls and agentic workloads burn more tokens, cost becomes a variable to manage rather than a fixed fee to forget — at the model layer and now at the developer-tool layer alike. The all-you-can-eat era is ending everywhere the bill is driven by inference, which by 2026 is everywhere.
Spectrum AI Lab — AI coding tools pricing 2026: Copilot vs Claude Code vs Codex vs Cursor → · Marissa Harcourt — The best AI coding tools in July 2026 →