Bolt Forge Ships a Vibe-Coding Agent That Runs on Open-Source Models at 50× Usage

StackBlitz has handed vibe coders a way to build flat out for a month without staring at a token counter. Bolt Forge, the company's new open-model coding agent, launched as a research preview on September 14, took the top spot on Product Hunt on the 18th, and promises individual Pro subscribers up to 50 times more usage at no extra cost through October 14. The headline is the 50x. The more interesting story is that one of the most popular full-stack builders on the internet has just built a lane that routes around OpenAI and Anthropic pricing entirely, and it is asking you to help train the model it runs on in exchange.

What Bolt Forge actually is

Bolt Forge is not a new editor. It is a model choice inside Bolt.new, the browser-based full-stack builder from StackBlitz. Where the standard Bolt experience runs on proprietary frontier models, Forge swaps in open-weight models that StackBlitz runs on its own reserved hardware. Because the company is not paying per-token API rates to a closed lab, it can hand out far more usage for the same monthly subscription.

The pitch is straightforward. Bolt's individual Pro plan costs $25 a month billed yearly. During the preview, Forge adds up to 50 times more usage on top of that, no access code required, no separate billing line. For a solo builder or a small agency that burns through Claude or GPT credits by lunchtime, that number is hard to ignore.

I should be precise about what 50x means here, because StackBlitz is. It is 50x more Forge usage on the Pro plan, not 50x more usage on the flagship models. You only get the extra headroom when you build inside Forge. That is the deal, and it is a real one rather than a fine-print trick.

Which open models does it run?

Forge routes work across several open-weight models. The defaults are GLM 5.3 Flash and GLM 5.3, and StackBlitz lists Kimi K3 and DeepSeek v4 Pro as experimental options you can switch into. If those names sound familiar, they should: they are the same open-weight models that have been driving the price collapse across the inference market all year.

That is the point. Bolt is treating these as a portfolio rather than a single model, and it says newly supported open models will land in Forge first. What you are really looking at is a routing layer sitting on top of whichever open-weight model is cheapest and strongest this month. That is a different business model from the closed labs, and it is worth understanding because it changes who sets the price of your build.

Under the hood, StackBlitz runs Forge on WebContainers, the technology that executes each project's development environment in the browser. The combination matters. Running the environment client-side and the models on reserved hardware is how the company gets its cost per build down far enough to offer 50x without losing its shirt.

So is the 50x real or marketing?

Both, honestly. The usage is real: build inside Forge and you get up to 50 times the allowance on your Pro plan through October 14. The catch is what you give up for it, and StackBlitz is unusually candid about that part.

Two things to know. First, there is a quality trade. Bolt's own blog concedes roughly nine points off its internal evaluation versus the flagship experience, while also claiming the open models reach around 91 percent on its test. Nine points is not nothing, and for a task where a model silently drops a subtle bug, a few points can mean more time debugging. Second, and this is the bigger one, the preview is a data flywheel. Forge is collecting real software-building traces that feed the training of an open-weight model with Arcee AI, the US open-model lab behind the trillion-parameter push. Your sessions are helping train the next model.

There is a privacy dimension here that builders should read before opting in. Bolt has been clear that Forge usage feeds the training effort, and the terms around what gets shared are the kind of thing you should check if you are building anything remotely sensitive. The 50x is a discount, but you are also the product in the literal sense.

Why this matters for the whole market

For the last eighteen months, vibe-coding tools have been quietly hostage to two companies' pricing. OpenAI and Anthropic set the token rate, and every builder layered on top of them inherited it. Bolt Forge is one of the first big mainstream builders to say out loud that it can run the whole experience on open weights and pass the savings on as a 50x usage bump.

That is a real moment. It does not mean open models beat frontier models today; they do not on the hardest reasoning tasks. It means the economics have shifted enough that a serious product is willing to ship open weights as the default experience, not a curiosity tucked behind a settings toggle.

I would not overstate it. The closed labs are not going anywhere, and most serious builders will keep paying for their frontier models when the task actually needs them. What Forge signals is that the middle of the market, the long tail of apps that do not need a frontier model, now has somewhere cheaper to go.

What I would watch before treating this as permanent

The preview ends October 14, and what happens after is the question that actually matters. If Forge becomes a permanent tier at anything like this usage ratio, it puts real pressure on every other builder still paying full API freight. If it quietly shrinks back to normal allowances once the preview closes, then this was a marketing campaign and a data-collection drive wearing a product-launch costume. Both readings are on the table right now.

I would also watch the quality gap in practice, not in a benchmark table. The builders I talk to care less about a nine-point eval difference and more about whether the agent can hold a multi-file refactor together at 2am. That is the test that matters, and no amount of free usage can paper over it.

The takeaway

Bolt Forge is worth trying for one specific reason: it is the clearest signal yet that open-weight models are good enough to be the default in a mainstream full-stack builder, and that the per-token tax from the closed labs is optional. Enable it on a throwaway project, not something with real customer data, and see whether 50 times the usage at nine points off is a trade you would make every day. My guess is that for a lot of scrappy builders, it is. And if the closed labs were not already nervous about pricing power, they should be now.