What happened

Bolt.new announced Bolt Forge, a new agent inside its AI app builder that launches September 14, 2026, as a research preview. It’s the browser-based tool where you describe an app in plain English and it writes and runs the code for you. BuilderWithin covered the same tool last week over its click-and-change Visual Edits toolbar, and this is a different kind of change: not a new editing feature, but a new choice about which model builds your app and what happens to your sessions.

Forge sits in the agent picker next to the Standard and Max agents Bolt users already have. The difference is what it runs on. Standard and Max use Bolt’s regular paid models. Forge runs only on open-weight models, meaning models whose underlying trained files are published for anyone to use or build on, rather than kept private by one company. By default Forge uses a model called GLM 5.3 Flash, with GLM 5.3 available alongside it, and Kimi K3 and DeepSeek v4 Pro offered as experimental options.

The trade Bolt is offering: every individual Pro plan gets up to 50 times more Forge usage than normal, at no extra cost, through October 14, 2026. In exchange, Bolt asks you to let your Forge sessions help train future open-weight models, through a partnership with Arcee AI, a company that builds open AI models. Bolt says this training is always opt-in: a consent screen appears every time you switch into Forge, spelling out the trade before you proceed, and sessions are anonymized with secrets stripped before they’re used. Standard and Max stay exactly as they are: Bolt says it doesn’t train on anything you build outside Forge.

On Bolt’s own quality benchmark, called the Bolt Build Index, Forge’s open models score 92.2 against 101.0 for Bolt’s top paid model, meaning Forge’s best output lands at roughly 91% of what the top paid model produces on that measure.

Why it matters

This is a real decision, not a free upgrade. Fifty times more usage sounds like an easy yes until you weigh what’s on the other side of it: every session you accept the consent screen for in Forge becomes training material for models other companies and developers can also use. If any of your work involves business logic, product ideas, or client work you’d rather not have influencing a model anyone else can download and run, that’s worth pausing on before you flip the switch.

The quality gap is smaller than the usage gap is large. A 9-point difference on Bolt’s own 101-point scale, roughly 91% of the top model’s score, is a real but modest drop, not a downgrade to something obviously worse. For a lot of everyday building, that gap may not be noticeable in the finished app. But if you’re already pushing what the top model can do, like intricate logic or an app that leans hard on one especially difficult feature, the gap is more likely to show up.

The bigger builder-facing signal is that Bolt needed a new incentive to get people using open models at all. Pro users weren’t reaching for open-weight models on their own merits before this offer existed, so Bolt built a usage allowance big enough to make trying them worth the trade.

Who should care

This matters most if you’re on a Bolt Pro plan and regularly bump against your usage limit, since Forge is aimed squarely at that friction. It’s also worth watching if you’re curious how open-weight models compare to paid ones for real app-building work, since Forge is a low-cost way to find out on your own project instead of taking a benchmark’s word for it.

It matters less if what you’re building involves information you wouldn’t want shaping a model other companies can use, or if you’re not running into Bolt’s usage limits in the first place. In either case, staying on Standard or Max costs you nothing you weren’t already paying.

What builders should do next

Before you switch into Forge for the first time, read the consent screen fully instead of clicking through it. It’s the one place Bolt tells you exactly what a given session will and won’t be used for, and it appears every time you switch in, not just once.

If you want to actually test whether Forge holds up for your work, pick one real feature you’d normally build with Standard or Max, something with a few moving parts, not a single button. Build it once in Forge and once in Standard or Max, using the same prompt for both. Compare four things: how many follow-up corrections each version needed before it worked the way you wanted, how much of your usage allowance each run consumed, how long each took start to finish, and whether either version handled an edge case you deliberately threw at it, like an empty form submission or an unusual input. Bolt’s 92.2-versus-101.0 benchmark score is Bolt’s own measure across its own test set, not a guarantee for your specific app, so this is the only way to know if the gap matters for what you’re actually shipping. If it doesn’t, the extra usage is close to free. If it does, save Forge for the lower-stakes parts of your build.

End of article