Bolt is giving developers 50x more compute. But there’s a catch.
The New Stack Amanda Caswell
Bolt is offering developers up to 50x more coding-model usage. The catch: they have to opt in to share anonymized session data for training.
Based on reporting by The New Stack, Amanda Caswell — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Bolt.new is trying a simple trade: more compute for more data. On Monday, StackBlitz’s browser-based coding platform launched Forge, a research preview for individual Pro subscribers that promises up to 50 times more usage of open weight coding models through October 14.
That extra headroom comes with a clear condition. Developers must explicitly opt in, and their sessions can be used for model training in anonymized form, including prompts, source code, fix traces, and conversations with the coding agent. Bolt says the data will feed a project with Arcee AI aimed at training a trillion-parameter-class open-weight model, with the first run slated to begin in October and the resulting weights eventually released publicly.
The pitch makes a certain amount of sense if you’ve watched coding agents chew through tokens on messy, multi-step tasks. Public repositories show finished code, not the dead ends, retries, and repairs that happened on the way there. Forge is betting that those messy trajectories are exactly what future agents need, especially as they work across large codebases and inherit code from other agents.
Bolt is also using its own infrastructure to make the math work. The agent currently runs on GLM 5.3 Flash and GLM 5.3, with Kimi K3 and DeepSeek v4 Pro available as experimental options. And because StackBlitz’s WebContainers keep projects isolated in the browser rather than on Bolt’s servers, the company already has one cost advantage before it even starts talking about model training.
The company says its open models scored 92.2 on its internal Bolt Build Index, versus 101.0 for its top paid model, or about 91% of the top score. That’s an internal benchmark, not a universal verdict. But it does explain the shape of the deal: Bolt gets a cheaper way to keep pushing open models, Arcee gets real developer traces, and developers get a lot more runway before the token meter starts screaming.
My take — AI-written commentary, not fact-checked reporting
This is the cleanest version of the new AI bargain: pay with your code, or pay with your wallet. Bolt is being unusually direct about it, which is refreshing in a sector that usually wraps data extraction in a privacy policy and a smile. The real tell is that open-model companies keep circling the same prize — real workflow traces — because synthetic neatness only gets them so far.
Read more about this at: The New Stack