Bolt.new is making a simple bet: developers should be able to experiment more without paying frontier-model prices for every draft.
On September 14, 2026, the StackBlitz-built AI app builder launched Bolt Forge, a new agent powered entirely by open-weight models.
For individual Pro users, the headline offer is aggressive: up to 50× more usage during a one-month research preview.
But there is a trade.
Users who opt into Forge agree to share anonymized build sessions so future open models can learn from something the public web rarely captures: how software actually gets built, broken, repaired, and refined inside a live development environment.
This is more than another model option in a dropdown.
It is a different economic model for AI coding.
Anyone who has spent serious time with an AI coding tool knows where the credits disappear.
It is rarely just the final polished prompt.
It is the messy middle:
The abandoned ideas.
The UI experiments.
The broken database connection.
The “try it another way” loop.
The feature you build, hate, delete, and rebuild.
That experimentation can consume the same premium-model credits as production work.
Bolt’s argument is straightforward: price has started shaping how freely people experiment with AI builders.
Forge is designed to remove some of that friction.
The idea is not that every task needs the best proprietary model available.
Some work simply needs cheap, plentiful iteration.
Forge is not one new model.
It is a third agent mode sitting alongside Bolt’s existing Standard and Max options.
Users choose Forge from the agent picker, select an available open model, accept the consent screen, and start building.
Switch back to Standard or Max and those workflows remain separate. Bolt says work completed outside Forge is not used for Forge training.
The launch lineup includes:
The latter two are positioned as more experimental and can consume the Forge allocation faster.
Bolt also says new frontier open models will increasingly appear in Forge first, effectively turning it into a testing ground for the latest open-weight coding systems.
There are limitations.
PDF support is not available yet, and Bolt is unusually clear that Forge should still be treated as experimental. The company recommends duplicating important projects before moving them into Forge and keeping complex production work on Standard or Max.
That caution is backed by Bolt’s own numbers.
On its internal Build Index, which measures how well models complete real Bolt projects, Forge’s open models scored 92.2, compared with 101.0 for the strongest paid model.
That puts Forge at roughly 91% of Claude Opus 5’s score as of September 2026.
Useful enough to matter.
Not good enough to declare premium agents dead.
The research preview runs from September 14 through October 14, 2026.
During that period, every individual Pro subscriber receives a large Forge allocation at no additional cost.
The system is intentionally simple:
Once the Forge bar is exhausted, users can continue building with Standard.
There is no surprise overage bill.
After October 14, Bolt expects Forge to remain available as an experimental environment, but the special 50× allocation and the initial training-data window are scheduled to end.
Teams and Enterprise accounts are excluded entirely.
No Forge.
No training data collected from those workspaces.
This is the most important part of the launch.
Forge is not simply free extra compute.
It is an exchange.
Every time a user switches into Forge, participation is opt-in. If they agree, Bolt can share anonymized prompts, code, and fix traces from those sessions.
Those fix traces may be the most valuable part.
They capture the sequence of things that goes wrong while software is being built:
an error appears,
the agent edits the code,
the fix fails,
another approach is attempted,
and eventually the application works.
That kind of data is difficult to scrape from GitHub or the open web because it captures the process, not just the finished artifact.
Bolt says secrets and personal information are removed before data leaves its infrastructure.
Its training partner is Arcee AI, the U.S.-based lab behind the Apache 2.0-licensed Trinity model family.
The initial Forge dataset is expected to feed training of a trillion-parameter-class open-weight model, with Bolt and Arcee saying the resulting model weights will eventually be released publicly.
There is one important caveat.
If someone later stops using Forge, data that has already been used in training cannot simply be removed from the trained model.
So the bargain is real:
You get substantially more room to build. Open models get access to the traces of how you build.
Giving users 50× more experimentation sounds expensive.
Bolt says several changes in AI infrastructure make it possible.
First, open-model inference is cheaper than relying entirely on frontier proprietary APIs.
Second, Forge runs on reserved hardware, reducing the cost associated with heavily marked-up per-request inference.
Third, Bolt still relies on WebContainers, its browser-based runtime technology.
That means much of the actual application execution happens locally in the browser rather than inside an expensive cloud development environment dedicated to every user session.
Put those pieces together and the economics begin to look very different.
As inference prices fall, it becomes increasingly realistic to give users dramatically more attempts without multiplying costs by the same amount.
The remaining question is quality.
Fifty times more drafts only matters if enough of those drafts are useful.
Forge also supports a second experiment: Bolt Lite.
The new plan costs $9 per month and is designed around Forge rather than Bolt’s premium agents.
It is aimed at students, hobbyists, and people building smaller weekend projects.
Lite includes the core Bolt builder, Bolt’s built-in database, collaboration tools, and publishing through a bolt.host URL.
But it is intentionally stripped down.
Users do not get Standard or Max, AI image generation, security scans, Supabase integration, custom domains, additional usage purchases, or the ability to remove Bolt branding.
Early access is scheduled to begin in waves around September 21, with sign-ups closing on October 14.
Anyone already subscribed to Lite at that point keeps the $9 monthly price.
This may be one of the more consequential parts of the experiment.
Forge is not only about giving Pro users more room to iterate.
It also gives Bolt a cheaper model stack capable of supporting a much lower-priced plan.
The most sensible workflow is probably not:
Replace Max with DeepSeek and hope for the best.
Forge makes more sense as the experimentation layer.
Use it to:
Then move back to Standard or Max when the work becomes more consequential:
That hybrid workflow is arguably the entire point.
Forge is the workshop. Premium agents remain the finishing tools.
The launch offer is easy to like.
The harder questions come after the research preview ends.
Three things are worth watching.
A 91% internal benchmark score is encouraging.
But the better test will be what happens when thousands of people throw messy SaaS ideas, abandoned side projects, half-working databases, and unpredictable requirements at Forge.
Benchmarks measure capability.
Real builders measure frustration.
The economics become much less interesting if the huge usage advantage disappears immediately after the research period.
If Forge turns into a lightly discounted experimental mode, much of the excitement fades.
If Bolt can maintain dramatically cheaper iteration — especially through Lite — it could meaningfully expand who can afford to build complete applications with AI.
The most ambitious promise is not the extra usage.
It is the dataset-to-model loop.
Bolt and Arcee are effectively saying:
Let people build more cheaply, use those build traces to train better open coding models, release those models publicly, then use the improved models to make building cheaper again.
That is a compelling flywheel.
But the proof will be downloadable weights.
Forge is ultimately Bolt arguing that drafting software should not cost the same as finishing software.
Most creative work involves far more experimentation than final output.
AI coding is no different.
If cheaper open models are already good enough for brainstorming, scaffolding, debugging, and early iteration, there is little reason every exploratory prompt needs to flow through the most expensive proprietary agent.
And if the messy work of building applications produces valuable training data, Bolt may be able to subsidize that experimentation while simultaneously improving the models powering it.
That is the interesting part of Forge.
It is not simply cheaper AI coding.
It is an attempt to create a loop where:
more usage creates better training data, better training data creates stronger open models, and stronger open models make more usage affordable.
If that loop works, Bolt may have found a way to make experimentation dramatically cheaper without abandoning premium models where they still matter.
If it does not, Forge may end up looking like a generous month of extra tokens attached to a data-sharing program.
For now, the experiment is live.
Individual Pro users can open Bolt’s agent picker, choose Forge, and decide whether the trade is worth it.
