The end of all-you-can-eat

Flat-rate coding assistants were priced for autocomplete. Agent workloads do not distribute like autocomplete, and the pricing is now catching up in public.

Cursor split team usage into two pools and added a $120 Premium seat. Claude Code's $2/$10 introductory API rates expire on 31 August. Copilot already moved every plan to usage-based billing.

Two pools is an admission about margin

Cursor pays market rates for Claude, GPT and Gemini calls and pays close to nothing for its own Composer models. A single blended allowance meant heavy third-party users were subsidised by everyone else — invisible, unsustainable, and impossible to steer.

Separating them makes the subsidy explicit. It also, not incidentally, gives Cursor a lever to push usage toward its own models.

The distribution broke the model

Autocomplete usage is roughly uniform across a team. Agent usage is not: a minority running long parallel loops consume most of the tokens. Flat per-seat pricing across that distribution loses money on exactly the users who like the product most.

Side Chats — parallel agent conversations — are the feature that made the pricing change unavoidable. Ship the thing that lets one developer run five loops at once and the old rate card stops working the same week.

Where this lands

Metered pricing on every tool turns model routing into a cost decision rather than a preference. Which is precisely the pressure that makes a capable local model interesting to finance departments who have never had an opinion about weights before.

The introductory-rate cliff on 31 August is the part to plan for now. Teams have spent months building workflows whose unit economics assume a number that expires in three weeks, against a replacement nobody has published.

Spectrum AI Lab — AI Coding Tools Pricing 2026: Copilot vs Claude Code vs Codex vs Cursor → · Developers Digest — AI Coding Tools Pricing Comparison 2026 →