Bolt Forge: Models, 50X Usage, Pricing and How the Research Preview Works

01

Overview

Bolt Forge is a new research-preview agent inside Bolt.new that is built around open-weight AI models rather than Bolt’s usual premium model stack.

Bolt launched Forge on September 14, 2026. It appears as a third agent option alongside Standard and Max for individual Pro users. The core trade is unusual: Bolt gives Forge users a much larger usage allocation during the preview, while users who enter Forge explicitly opt in to let anonymized build sessions contribute to training future open-weight models with Arcee AI.

That makes Forge more than a cheaper model picker. It is simultaneously an app-building agent, an open-model testing environment and a data-sharing experiment intended to improve future coding models.

02

What is Bolt Forge?

Bolt Forge is an agent inside Bolt.new, the AI app and website builder from StackBlitz. It is not a single model. Instead, Forge can run multiple open-weight models inside Bolt’s existing building environment.

At launch, Bolt lists GLM 5.3 Flash as the default model, with GLM 5.3 also available. Kimi K3 and DeepSeek v4 Pro are offered as experimental options.

Bolt says the lineup will change as new open models become available. That means users should treat the current model list as a launch snapshot rather than a permanent catalog.

03

When did Bolt Forge launch?

Bolt launched the Forge research preview on September 14, 2026.

The first preview window runs through October 14, 2026. Bolt says Forge will continue after that date as an experimental open-model lab, but the temporary launch allocation and first Arcee AI training window are tied to the September 14–October 14 period.

04

Which models does Bolt Forge use?

The launch lineup includes:

• GLM 5.3 Flash — the default Forge model at launch.
• GLM 5.3 — available alongside Flash.
• Kimi K3 — experimental inside Forge.
• DeepSeek v4 Pro — experimental inside Forge.

Bolt warns that these models do not consume the Forge allocation at exactly the same rate. Kimi K3 and DeepSeek v4 Pro use the available Forge allowance faster than the GLM options.

The important distinction is that Bolt Forge is an orchestration/product layer around these models. Searching for “Bolt Forge model” should not be interpreted as meaning Bolt released its own foundation model.

05

What does “50X more usage” mean?

During the research preview, Bolt says every individual Pro plan receives up to 50X more Forge usage at no extra cost through October 14, 2026.

Forge usage is kept separate from Standard and Max usage. Instead of a daily allowance, Forge uses one monthly usage bar that resets on the account’s renewal date. Bolt says there are no daily limits and no automatic overage charges. When the Forge bar reaches 100%, the user is moved back to Standard rather than billed for additional Forge usage.

The “50X” figure is a launch allocation claim from Bolt. It should not be interpreted as a permanent pricing promise, and Bolt has not yet published the exact post-preview allocation that will apply after October 14.

06

How much does Bolt Forge cost?

Forge is included for individual Pro users during the preview rather than sold as a separate add-on.

Bolt’s current pricing page lists Pro at $25 per month, while the Forge launch article says the temporary Forge allocation is included at no additional charge for eligible individual Pro users through October 14, 2026.

Bolt also announced Bolt Lite as another access path. Lite seats are distributed in waves, with the first seats expected by September 21, 2026. Bolt says users who are on Lite when sign-ups close on October 14 can keep that plan at $9 per month.

Those numbers are time-sensitive. Anyone evaluating Forge should check Bolt’s live pricing and access pages before buying because the research preview is explicitly temporary.

07

How does the data-sharing model work?

Forge requires an explicit opt-in whenever a user switches into it.

According to Bolt, Forge sessions can include prompts, code and the fix traces generated during builds. Bolt says it strips secrets and personal information and anonymizes the data before it leaves Bolt infrastructure. The resulting data can then be transferred to Arcee AI under a data-processing agreement and used to help train open-weight models.

Standard and Max are separate. Bolt says it does not use work from those agents for Forge training simply because the same account uses Bolt.

Teams and Enterprise workspaces are excluded from Forge and from Forge data collection at launch.

There is one important permanence caveat: switching back to Standard or Max stops new Forge sharing, but data already used in model training cannot be pulled back out of the models it helped create.

08

Why is Bolt working with Arcee AI?

Bolt says Forge sessions will help Arcee AI train future open-weight coding models using real software-building workflows.

The partnership is notable because the training data is not simply public code scraped from repositories. Bolt is trying to capture the sequence of how people actually build: creating a project, making edits, connecting services, encountering errors and fixing them.

Bolt says the first training run begins in October 2026 and that the resulting model weights will be published. That is a forward-looking vendor commitment and should be rechecked when the training run and resulting model releases actually occur.

09

How good is Bolt Forge?

Bolt reports that Forge’s open-model lineup scored 92.2 on its internal Bolt Build Index, compared with 101.0 for the top paid model in the benchmark, which Bolt identifies as Claude Opus 5. Bolt translates that to roughly 91% of the top score.

This is useful product information, but it is not an independent benchmark. Bolt designed the benchmark, ran the tests and published the result. It should therefore be described as a Bolt-run evaluation rather than proof that Forge is 91% as capable as a premium model across all coding tasks.

Real-world performance will also vary by model, project type, prompt quality and how much context a build requires.

10

Is Bolt Forge good for production apps?

Bolt itself recommends caution.

The company labels Forge’s open models experimental and advises users to duplicate serious projects before switching them into Forge. Bolt also suggests keeping complex production work in Standard or Max.

That makes Forge especially attractive for brainstorming, prototypes, MVPs, experiments and iterative builds where high usage limits matter more than always using the most capable premium model.

For production-critical work, users need to evaluate model quality, privacy requirements and the cost of mistakes rather than choosing Forge only because the usage allowance is larger.

11

What are the current limitations?

At launch, Forge has several important limitations.

First, the model lineup is experimental and may change.

Second, Forge does not currently support PDF uploads.

Third, Teams and Enterprise workspaces are not eligible for Forge.

Fourth, the temporary 50X allocation runs only through October 14, 2026. Bolt has not yet published a permanent post-preview usage policy.

Fifth, Bolt’s quality claims are based on its own Build Index rather than an independent benchmark.

12

Bolt Forge vs Standard and Max

The main difference is not simply model quality.

Forge uses open-weight models and requires an explicit opt-in to the Forge data-sharing/training arrangement. It is designed to provide much more experimental usage during the launch preview.

Standard and Max use Bolt’s regular premium-agent stack and are not part of Forge training. They remain the safer default for users who do not want to share Forge sessions or who are working on more demanding production projects.

That means the choice is a three-way trade between cost/usage, model capability and data-sharing preferences.

13

Who should use Bolt Forge?

Forge is a strong fit for individual builders who:

• already use Bolt Pro;
• want more room for prototyping and iteration;
• are comfortable opting in to anonymized training-data sharing;
• want to experiment with open-weight coding models inside a polished app-building environment;
• do not need Teams or Enterprise workspace support.

It is a weaker fit for users handling highly sensitive code, teams that cannot accept the data-sharing model, or projects where experimental model behavior creates unacceptable production risk.

14

Is Bolt Forge free?

Not as a standalone public free product at launch. Forge is included with individual Pro plans during the preview. Bolt Lite is a lower-cost access route, but access is being distributed in waves rather than opened universally on day one.

15

Does Bolt Forge train on your code?

Only if you switch into Forge and accept the opt-in flow, according to Bolt. Forge sessions can be used for training after Bolt says it removes secrets and personal information and anonymizes the data. Standard and Max sessions are not automatically included in that training arrangement.

16

What happens after October 14, 2026?

Bolt says Forge will continue as an experimental open-model lab. What is not yet defined is the permanent usage allocation and commercial structure after the research-preview promotion ends.

17

Bottom line

Bolt Forge is one of the more interesting AI-builder launches because it changes both the economics and the data relationship between users and the models they build with.

The short-term pitch is simple: individual Pro users get a much larger experimental allocation for open-model app building. The longer-term experiment is more consequential: opted-in real-world build sessions are being turned into training data for future open-weight coding models.

That makes Forge worth watching even for people who do not use Bolt.new. If the approach works, other AI coding and website-building platforms may increasingly offer cheaper open-model tiers in exchange for carefully permissioned training data.