Bolt.new token-based margin risk

The core issue is that Bolt sells software like a subscription, but buys intelligence like a metered utility. Every prompt, file edit, and retry burns tokens, so gross margin gets squeezed precisely when users attempt larger, more valuable projects. That is a different shape of software business than Cursor or Copilot, which can smooth costs with smaller in house models and flatter pricing. For Bolt, power usage and margin pressure rise together.

• Bolt’s usage pattern is expensive by design. It is not just generating one answer, it is reading the existing project, understanding multiple files, writing code, and revising it. Bolt’s own token guide says most usage comes from reading and syncing project files, which means bigger apps naturally get more costly to serve.
  2 bolt 5 sacra
• The comparison set matters. Bolt has been priced around $20 per month with token limits, while Cursor and Copilot have been able to offer flatter $10 to $20 subscriptions because they mix frontier models with cheaper fine tuned models for smaller tasks. Bolt is more directly exposed to frontier model inference costs.
  1 sacra 3 sacra
• This also explains why simple prototyping is a better fit than long lived, iterative development. In the broader workflow, users often start a project in Bolt or Lovable, then move the code into Cursor or another local IDE for heavier editing. The handoff happens where token burn and platform constraints start compounding.
  3 sacra 4 sacra

The likely path forward is more packaging around this cost curve, with tighter limits, team plans, top ups, and selective routing to cheaper models for low stakes tasks. The winners in AI coding will be the ones that turn expensive model calls into predictable software margins, while still feeling magical on the first prompt.