Bolt Forge

Inside Bolt.new, Forge joins Standard and Max as a third agent, and every model behind it is open source. The company recommends GLM 5.3 Flash for most work, with GLM 5.3 also available and Kimi K3 and DeepSeek v4 Pro offered as experimental options that use up an allowance faster. Bolt announced it on 14 September 2026 as a research preview lasting until 14 October, followed by a broader launch post five days afterwards.

The trade it offers

Open models cost less to serve, so Forge comes with far more usage than the other agents: as much as fifty times more for a Pro subscriber, at no added cost during the preview. The price is consent: sessions become training data. Bolt lists what it gathers as prompts, code, project files and configuration, the tool calls it makes, edit histories and the fix traces it generates. After de-identification the material goes to a partner that trains open-weight models, and Bolt leaves out sessions that start in the EEA, the UK or Switzerland.

Who can get it

Pro subscribers get Forge automatically. It is also part of Bolt Lite, a new standalone plan that had a waitlist when this was written. Teams and Enterprise plans do not include it.

What happens after the preview

What follows is uneven, which is worth checking before signing up. Lite subscribers keep Forge with no end date, whereas Pro subscribers lose it after 14 October and retain Standard and Max. More Forge usage cannot be purchased, allowances reset on the renewal date and there is no daily cap. PDF uploads do not work yet.

Capability, in Bolt's own numbers

Bolt says Standard and Max remain its most capable agents. It publishes its own comparison: on Bolt's Build Index Forge scores 92.2, while the best paid model in the product scores 101.0. Both figures are Bolt's own measurement on its own index.