Baseten is launching three products this week that stretch it from hosting AI models to governing, training and housing them. The company sells inference (the work of running a finished model to answer a request), and Amir Haghighat laid out the new lineup in a post on the Baseten blog on 3 November.
Grid comes first. Baseten says it opens as a public preview the day after that post, and it is a gateway, router and set of guardrails for what companies spend on AI coding tools. The pitch is about sprawl: Baseten says coding agents have pushed bills up sharply, usage is scattered across models, vendors and tools, and developers reach for the smartest model on offer instead of the cheapest one that does the job. It adds that open-weight models, the kind anyone can download and run, now match top proprietary models on coding at a lower running cost.
Grid has three parts. The gateway puts every model and every dollar in one view. The router picks a model for each task, and Baseten says it also tries to keep a long agent session on a “warm” cache so repeat work gets cheaper as it goes. The guardrails set company-wide rules on who may use which model and what the models may do. It is built to work with Pi, OpenCode, Codex and Claude Code. Baseten says Opendoor already sends every AI request in the company through Grid without changing the tools its developers use.
The post gives no figure for how much Grid saves. That gap matters, because cost control is the whole sales case.
Sandboxes are the second launch. A sandbox is a sealed virtual computer where an AI agent can run code, install packages and make a mess without touching anything real. Baseten recently acquired Blaxel, which built the technology, and did not say in the post what it paid. Baseten describes the sandboxes as stateful, pausable and resumable, isolated at the level of a full virtual machine, and able to grow to thousands of copies within seconds. They move from private to public preview on Thursday. Because they will sit in the same data centers as the models, Baseten argues the back-and-forth between an agent thinking and an agent running code gets faster and easier to monitor.
Loops is the third. Fine-tuning means taking an existing model and training it further on a company’s own data so it does one job well. Loops does this on open-weight models through a serverless service, meaning the customer never manages the machines. Baseten names Datadog, Harvey, Lovable and Parallel as teams it helps customise open models, and says Datadog trains on Loops. Its software kit is compatible with Tinker, and customers can ask for single-tenant capacity, meaning hardware not shared with other companies. A finished model can be served straight from Baseten’s dedicated inference. Loops has been in public preview for some time and becomes generally available on Friday. Baseten says customers always keep ownership of their trained weights.
That last promise is the one to read twice. The weights are the thing a company has paid to create, and a vendor that lets you walk away with them is easier to say yes to.
All three launches follow a single logic: Baseten wants the model, the training run and the agent’s workspace in one place, so a company has one bill, one uptime promise and one place to set rules. Baseten claims 99.99 percent uptime for the combined platform. That figure comes from the company and arrives without any independent measurement.
A team whose coding-agent bill keeps climbing has a cheap experiment available this week: send a slice of traffic through Grid’s preview and compare the cost of each finished task against the current default model.
Reported by the Baseten blog, in a post by Amir Haghighat published on 3 November 2026.
