Bolt Forge offers 50x usage for coding traces

AI app builder Bolt.new has launched Bolt Forge, an opt-in coding agent powered by open-weight models. Individual Pro subscribers can receive a Forge allowance up to 50 times their existing usage allocation at no added cost through October 14, 2026, when they consent to share prompts, code, and agent-generated fix traces for model training.

• Status: Research preview in Bolt’s agent picker alongside Standard and Max.
• Default model: GLM 5.3 Flash, with three other models available.
• Data shared: Prompts, code, and the sequence of errors, edits, and fixes produced during a session.
• Consent: Required each time a user switches into Forge.
• Availability: Individual accounts only; Teams and Enterprise workspaces are excluded.

Bolt is using the preview to collect real software-building traces for Arcee AI, a U.S. open-model lab. Builders receive more room for experiments and prototypes, while Arcee receives training examples that ordinary source-code snapshots rarely contain.

Open models reach 91% in Bolt’s test

Forge routes work among several open-weight models. GLM 5.3 Flash is the default, with GLM 5.3 also available. Kimi K3 and DeepSeek v4 Pro appear as experimental options. Bolt plans to place newly supported open models in Forge first.

Open-weight means a model’s learned parameters are available under a license. Source code, training datasets, and usage rights can still vary among models.

In Bolt’s September 2026 Build Index, the Forge lineup reached 91% of the score posted by Claude Opus 5, the highest-scoring paid model in that test. The result measures performance on real Bolt projects, according to the company, and leaves a nine-percentage-point gap between Forge and the leading paid option.

The Build Index is an internal, product-specific benchmark designed and run by Bolt. Independent evaluation would be needed to generalize the result to other development environments, languages, or agent workflows.

Forge remains experimental inside Bolt, and the current preview carries four operational limits:

• Duplicate a project before moving a consequential build into Forge.
• Keep complex, production-critical work in Standard or Max.
• Kimi K3 and DeepSeek v4 Pro consume the Forge allowance faster than the GLM models.
• PDF uploads are currently unsupported.

Consent follows every switch

Each switch into Forge opens a one-tap consent screen, keeping data sharing tied to the selected agent and current session. Standard and Max sessions remain outside the Forge training pipeline.

Bolt says an opted-in Forge session can contribute three categories of material:

• Prompts: The instructions and follow-up requests sent to the agent.
• Code: Source created, edited, or supplied during the session.
• Fix traces: The sequence of generated edits, errors, retries, and repairs that leads toward a working build.

Before data leaves Bolt’s infrastructure, the company says its pipeline removes secrets and anonymizes personal information by default. Bolt validates those filters with seeded test data. The remaining prompts, code, and traces continue to Arcee for training.

Switching back to Standard or Max stops future Forge sharing. Data already incorporated into a training run cannot be removed from the resulting model weights. Teams and Enterprise workspaces have no Forge access and are excluded from this collection program.

Arcee is the developer of the Apache 2.0-licensed Trinity model family. The partnership aims to train an open-weight model in the trillion-parameter class, with the first training run scheduled for October 2026. Parameter count describes model size; capability still requires evaluation.

The published description promises that the resulting weights will be released openly. It does not specify the new model’s license, the retention period for raw sessions, or a process for deleting contributed data before training begins.

Reserved compute buys more headroom

Forge’s larger allowance relies on reserved hardware and browser-based project execution. Reserved capacity gives Bolt more predictable inference costs than per-request provider billing. WebContainers, the StackBlitz technology underlying Bolt, runs each project’s development environment in the browser and reduces the server capacity required for application builds.

Forge metering isolates the preview allowance from existing paid-agent credits:

• A single monthly bar tracks Forge consumption.
• No daily usage caps apply.
• At 100%, Bolt moves the user back to Standard without an overage charge.
• Forge activity does not consume Standard or Max allocations.

A $9 route into Forge

Bolt is also introducing Bolt Lite, a lower-priced plan aimed at students and side projects.

Bolt Free and Lite plan comparison

Feature	Free	Lite
Monthly price	$0	$9
Relative AI usage	1x	50x
Agent	Standard	Forge
Positioning	Exploration	Side projects

Lite invitations will roll out in waves, and sign-ups close October 14, 2026. Subscribers admitted during that window keep the $9 price indefinitely. Pro costs $25 a month with annual billing and includes the same 50x Forge allowance through the preview window.

Build traces fund the discount

Software-building traces capture the process between a request and a completed project: prompts, proposed edits, compiler or runtime errors, retries, and final repairs. Public repositories usually preserve source snapshots and commit history, leaving much of that problem-solving sequence unavailable for model training.

Those traces can help Arcee train models to use development tools, recover from errors, and complete longer coding tasks. Bolt’s arrangement assigns that data an explicit price through additional usage and promises to publish the resulting model weights.

Developers and researchers can also use Forge to compare open models on the same Bolt projects. Useful measurements include output quality, repair-loop count, completion rate, elapsed time, and allowance consumption by model, all of which reveal project-level tradeoffs that an aggregate benchmark cannot show.