CoreWeave Forge Launches: $60/Month for Small Teams and Per-Millisecond Sandbox Billing

Bottom line: CoreWeave launched Forge on September 30, 2026. Forkast reports a $60-per-month plan for teams under 50 people, while serverless sandboxes are billed separately by the millisecond.

The diagram and the invoice are not the same object. CoreWeave’s Sept. 30 newsroom post introduces Forge as one connected environment for training, inference, evaluation, and agent development, open across models, frameworks, and other clouds. Blair Hayes at Forkast on Oct. 4 is the piece that puts numbers on a buyer: $0 for individual builders, $60 a month for teams under 50 employees, custom pricing after that, and a 30-day Pro trial. Serverless sandboxes, in that same Forkast read, use Kata VM isolation and bill by the millisecond. If you only remember the loop diagram, you will miss the line item that actually changes a budget.

This is a tool review of a bundle, not of a new model. CoreWeave was already selling the GPUs. Forge is the claim that the traces, the checkpoints, and the agent runs should sit on that same infrastructure instead of being handed across vendors. Solutions Review’s week-of-October-2 roundup describes the same shape in one paragraph: training, inference, observability, evaluation, curation, post-training, and agent development, combining Weights & Biases Models, OpenPipe, marimo notebooks, and CoreWeave infrastructure, with ARIA, Agent Lens, Sandboxes, a Registry, and serverless fine-tuning and reinforcement learning named as features. Three outlets, one product, and only one of them published a team price.

We have reviewed other control layers this fall. JetBrains Air is a control layer, not a new model. IBM Bob can be self-hosted, and you still bring the model. Goose lets you swap models and does not swap out the bill. Forge is a fourth shape. The landlord is trying to become the floor the agent runs on. That can be a real product. It can also be a rename of tools you already pay for, plus two new meters. The review is which sentence you can verify from the posts, not which diagram looks closed.

What Forge says it connects

Model and agent development, the newsroom post says, has been split across tools from separate vendors, none built to work together. The cost it names is specific: a production trace does not feed the next training run, a finished experiment does not inform the next evaluation, and every handoff loses a signal or costs more time. That is a workflow complaint. It is not a benchmark.

The components it then lists are how you test whether the complaint got a product or a sitemap. Agent Lens is the observability and continuous-improvement tool. Sandboxes are generally available, a fresh isolated environment for every run, whether that run is agent tool use, reinforcement learning, or model evaluation, already connected to the rest of the loop so nothing extra has to be stood up. ARIA is generally available, described as a coding agent that analyzes large-scale experiment data and agent observability data, surfaces what drove a change, and proposes the next experiments and the agent worth running, with evidence attached. Weights & Biases Models is the experiment tracker: tens of thousands of experiments, millions of metrics, visualizations, automated workflows. Notebooks are the developer surface that is supposed to live where the work runs, so a prototype carries into training, evaluation, and production instead of dying in a rewrite. Custom visualizations, CoreWeave says, are built on marimo. The Registry versions every model checkpoint and agent configuration in open, portable formats, with lineage, so the next attempt can start from any point in the history, on any stack. Post-training is a set of services: serverless reinforcement learning, serverless supervised fine-tuning, and model distillation.

Forkast compresses the corporate story. Forge was signaled at Fully Connected 2026 in San Francisco. The loop, in Hayes’s wording, is Run, Observe, Curate, Improve, Evaluate, on the same infrastructure that serves the GPUs. He says the platform consolidates Weights & Biases, OpenPipe, and Marimo, alongside newer services including model distillation, Notebooks, and ARIA, which he calls a research agent now generally available. The pieces he treats as structurally significant are Agent Lens, Sandboxes, and the Registry, because those are the ones that move a compute supplier into the layer where agents execute, get observed, and get versioned.

A notebook product and an experiment tracker are things a lab may already have, including from the companies CoreWeave has folded in. An isolated execution environment with a published isolation technology and a billing unit is a new meter. A registry with a portability claim is a lock-in argument, and lock-in arguments need the file format, not the slogan. CoreWeave says open, portable formats. “Open” without an extension is a slide. Ask for the extension before you treat the Registry as an exit door.

Solutions Review does not add a price or a format either. When a roundup and a newsroom post name the same feature set, the feature set is at least what the company wanted printed in the first week of October.

Agent Lens is a benchmark sentence

The number people will quote is the 20 percent. CoreWeave’s newsroom post says Agent Lens analyzes tens of millions of traces and automatically turns them into insights that drive proven fixes. It then says Agent Lens is detecting 20 percent more critical failures, and fixes issues at one-tenth the cost, based on a benchmark against a general-purpose frontier LLM. Forkast’s version is close and not identical. Hayes writes that Agent Lens traces every step, decision, and tool call a production agent makes, end to end, and that monitors score live traffic against baselines with human oversight, surfacing failures before they become patterns. He attributes the 20 percent failure-detection improvement and the roughly 10x lower cost to CoreWeave’s report, compared with a general-purpose frontier LLM benchmark.

Read those as the same claim wearing two coats. “One-tenth the cost” and “10x lower cost” are the same ratio if you mean the fix costs a tenth as much. They are not a price list. Neither post names the frontier model, the trace set, the definition of a critical failure, or who ran the benchmark. A vendor benchmark against an unnamed general-purpose model is a direction, not a procurement number.

Monitors score live traffic against baselines with human oversight. That is a product shape: not a fully closed loop, a scored stream that a person is still supposed to look at. If your reason for buying Agent Lens is to fire the on-call reviewer, the Oct. 4 description does not support that. It supports a tool that finds more of the failures the reviewer would otherwise miss, at a claimed lower cost than asking a frontier model to do the same pass. Those are different headcount stories.

ARIA sits next to this and is easy to confuse with it. Agent Lens watches production agents. ARIA, in the newsroom wording, analyzes experiment data and observability data and proposes the next experiments and which agent to run, with evidence. Forkast calls ARIA a research agent and says it is generally available. A coding agent that proposes experiments is not the same SKU as a tracer that scores live traffic. If a sales call blends them into “the AI that improves your AI,” split the sentence back apart before you sign. One proposes. One traces. The newsroom post marks both generally available. General availability is not “included in the $60 tier.” Forkast does not map features to tiers.

Two sandbox bills, not one

Sandboxes are the section where the tool stops being a diagram. CoreWeave says they are generally available, a fresh isolated environment per run, for tool use, reinforcement learning, or evaluation, already wired into the rest of the loop. Forkast adds the split that changes the invoice. The serverless variant uses Kata VM-based strong isolation and bills by the millisecond. The on-cluster variant runs inside a customer’s existing CoreWeave Kubernetes Service at no incremental cost.

Those are two products. If you are already on CoreWeave Kubernetes, the on-cluster line is the one to verify with your account team, because “no incremental cost” can mean the sandbox does not add a SKU, while the GPUs you already burn still cost what they cost. If you are not on that Kubernetes service, the millisecond meter is the one that will show up. Millisecond billing is precise and also easy to underestimate. An agent that retries tool calls in a loop does not feel expensive at the scale of one failed curl. It feels expensive at the scale of a night of retries. The post does not publish a per-millisecond rate. It publishes the unit. Ask for the rate before you treat “serverless” as the cheap path.

Kata is a concrete isolation claim, which is more than most agent-platform announcements give you. It is still not a security review. Forkast says strong isolation. If the sandbox is where a production agent runs tools against your data, the question for security is which network the Kata VM can reach, not whether the marketing word was “isolated.” That absence belongs in the review.

Forkast also places CoreWeave next to a different cloud’s announcement: it reports that DigitalOcean shipped Agent Droplets the following day, Oct. 1, in public preview, at $50 a month, with Firecracker microVM isolation and governed access to 16,000 tools. The comparison, as Forkast draws it, is architectural: a compute provider bundling agent execution, tool access, and governance, from the general-purpose cloud layer rather than the GPU layer. A $50 preview price and a $60 team price are not a bake-off. They are two numbers printed in the same week, for products that do not claim the same hardware. Use them to notice that “agent droplet” and “agent sandbox” became line items at the same time. Do not use them to declare a winner.

$60 is a team cap, not a usage cap

Forkast’s pricing paragraph is short, which is why it is useful. $0 for individual builders. $60 per month for teams under 50 employees. Custom enterprise pricing. A 30-day Pro trial. Early customers named: MasterClass and Canva. A $0 individual plan can be a real sandbox with a tiny quota, or a login that cannot launch the thing the keynote demoed. A $60 plan capped at teams under 50 is a headcount gate, not a GPU gate. Fifty employees is a company size. It says nothing about how many millisecond-billed runs those fifty people can start.

One is the $60, if you are under the headcount line and the trial convinced you the tier includes the sandbox you need. Three is the serverless meter, if you choose that variant, at a rate you have asked for in writing. Enterprise custom pricing is the bucket where all three go to a salesperson. MasterClass and Canva are logos, not case studies. Hayes names them as early customers. He does not say what they deployed, what they pay, or whether they use the millisecond sandbox or the on-cluster one. A logo is not a reference call. Treat it as a logo.

The post-training numbers are a fourth bill, and they are the closest CoreWeave comes to a comparative cost in its own newsroom post. Serverless reinforcement learning, it says, trains 1.4 times faster at 40 percent lower cost than a self-managed setup. Again: no hardware list, no self-managed baseline spec, no date on the measurement. A 40 percent claim against “self-managed” is the claim every managed service makes until you define self-managed. Those are different 40 percents. The newsroom post does not say. The honest review line is that they published a ratio and not the worksheet.

Weights & Biases and OpenPipe and marimo are the acquisition story sitting under the new names. Forkast says Forge consolidates them. Solutions Review says the platform combines Weights & Biases Models, OpenPipe post-training expertise, and marimo notebooks. If you already pay Weights & Biases, the question is whether Forge replaces that contract, discounts it, or adds CoreWeave on top. None of the three posts answer it. A bundle announcement that does not say what happens to the existing subscription is incomplete on purpose, or incomplete because the commercial team had not finished. Either way, do not cancel the old invoice because a newsroom post used the verb “combines.”

Questions to ask the vendor

Which of Agent Lens, ARIA, Sandboxes, Registry, and serverless post-training are in the $60 plan, and which start a different meter.

If the appendix does not exist, the 20 percent stays in the press post and out of the internal memo. CoreWeave’s newsroom post is still the right primary document for what the company says shipped on Sept. 30. Forkast is the right document for the price and the Kata meter, four days later. Solutions Review is a check that the feature list did not mutate in the roundup. Together they are enough to decide whether Forge is a platform you need a meeting about. They are not enough to migrate a production agent on a Friday.

The meeting is the product, for now. That is not an insult. It is what a GPU company looks like when it starts selling the loop around the GPU and publishes one team price, one billing unit, and a benchmark with the model name missing. Take the $60 seriously. Take the millisecond seriously. Leave the closed-loop adjectives on the slide.

What this means for you

When estimating cost, treat Forge as two line items: the fixed team subscription and variable sandbox usage. Ask CoreWeave which features sit in each tier and how sandbox time is metered before running production agent workloads on it.