Bolt.new gives Pro users 50 times more compute for sharing their sessions
The New Stack reports that Bolt.new launched a research preview called Forge, which gives individual Pro subscribers up to 50 times more usage on open-weight coding models through 14 October, provided they let their sessions train a new trillion-parameter model with Arcee AI.
Why it mattersA team paying for a coding-agent subscription now has to decide what its source code and its retry traces are worth as training data, because that data pays for the extra compute, and this is the first mainstream product to offer that trade.
The New Stack reports that Bolt.new, the browser-based coding platform from StackBlitz, launched a research preview on Monday called Forge, which gives individual Pro subscribers up to 50 times more usage of open-weight coding models until 14 October in exchange for training data from their sessions. Team and Enterprise accounts cannot participate, and the deal requires an explicit opt-in each time a developer switches into the Forge workspace.
What Bolt takes in exchange
Per The New Stack, developers who use Forge agree to share anonymised versions of their sessions, including the prompts they wrote, the source code the agent touched, the fix traces produced as they worked through problems, and the conversations with the coding agent itself. The data is used in a joint project with Arcee AI to train a trillion-parameter-class open-weight model, with the first training run scheduled to start in October. Bolt says the resulting weights will be released publicly.
The report says Bolt strips secrets, sensitive data and personal information before the sessions leave its infrastructure, and Arcee receives the data under a signed processing agreement. A developer can stop sharing new sessions by leaving Forge, but The New Stack says Bolt has confirmed that anything already used to train a model will stay in it.
How Bolt gets to 50 times
The 50x figure comes from switching the underlying models. According to The New Stack, Forge runs open-weight models on Bolt's own hardware rather than paying per token to a proprietary API. The agent currently uses GLM 5.3 Flash and GLM 5.3 by default, with Kimi K3 and DeepSeek v4 Pro available as experimental options. Bolt's WebContainers technology already runs the developer's project in the browser rather than on a server, which lowers the rest of the compute bill.
Bolt says the open-weight models score 92.2 on its internal Bolt Build Index, against 101.0 for its top paid model, which The New Stack calls 91 percent of the top score. The New Stack also notes plainly that the number only measures performance inside Bolt, and does not compare these models to anything else.
What a team should decide
A team choosing whether to switch into Forge is trading its retry traces for a token budget. The retry trace, the record of every path that did not work, is one of the more valuable things the coding assistants collect: it is what SpaceXAI trained Grok 4.6 on, according to earlier reporting from The New Stack, and it is what Arcee's own open-source agent harness NAC is built to use. A team's failed attempts are training data for whoever holds the transcript.
The migration cost is smaller than usual because there is no migration: Forge is a workspace inside the existing Bolt product, opened only after a consent screen. The decision is per developer per session. The 50x increase ends on 14 October, but Bolt says Forge itself will remain after the preview as a place to test open models.
Source
- Bolt is giving developers 50x more compute. But there's a catch., The New Stack, 15 September 2026.
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.
