Bolt.new Adds Forge, Offering 50x More AI Coding Usage for Opt-In Training Data

Developers on Bolt.new’s individual Pro plans can now get up to 50 times more AI coding-agent usage through October 14. The catch is deliberate: choosing the new Forge mode means opting in to share anonymized build sessions for open-weight model training.

Bolt launched Forge on September 14 as a research preview inside its agent picker. The mode runs open-source models only, including GLM 5.3 Flash by default, alongside GLM 5.3 and experimental Kimi K3 and DeepSeek v4 Pro options.

The offer turns build sessions into training material

The material is broader than finished code. Opted-in Forge sessions include prompts, code, and fix traces: the edits, errors, retries and repairs generated while trying to complete a project. Bolt sends the anonymized data to Arcee AI under a signed data processing agreement to help train open-weight models.

Consent is repeated, but not retroactive

A consent screen appears every time a user enters Forge. Declining leaves the user in Bolt’s usual agents, while Standard and Max sessions are not used for training. Switching back stops new sharing, but Bolt says material already used in training remains in the resulting model.

Bolt and Arcee are targeting a trillion-parameter-class open-weight model, with a first training run expected to begin in October. Bolt says the resulting weights will be published, making the preview a data-for-capacity exchange tied to a planned public model.

More room to experiment, with clear limits

Bolt says Forge’s models scored 92.2 on its internal Build Index, against 101 for Claude Opus 5—91% of the top score. That comparison covers real Bolt projects, according to the company, but it is not an independent measure of performance across coding environments or workflows.

How Forge is fenced off

• Forge usage has its own monthly meter, separate from Standard and Max, with no daily limits.
• When that allowance is exhausted, Bolt returns the user to Standard without an overage charge.
• Bolt advises duplicating serious projects before using Forge and keeping complex production work in Standard or Max.
• Teams and Enterprise workspaces are excluded from Forge and its data collection.

The exchange is visible rather than buried in a one-time setting: extra experimentation capacity in return for a record of how software gets made and fixed. Its practical boundary is equally clear—once a contribution has helped train a model, leaving Forge cannot pull it back out.