Beam free credits: None

Beam published $30 of free credit per month until early September 2026. It does not any more. On 2026-09-09 the pricing page carries per-second rates and nothing else: no credit figure, no “Free Credit, Every Month” line, and no plan tiers at all. The docs billing page states only that storage is free to 1TB. The withdrawal is recorded here rather than quietly deleted, because the $30 figure is still circulating in roundups and is the kind of number people budget against.

What the pricing page says now

The page was rebuilt around six tabs, Serverless, Sandboxes, On-demand, Clusters, Storage and BYOC, plus a calculator. The headline reads “Pay as you go, billed by the millisecond”, and the teaser line is “Sandboxes from $0.319 /hr · H100s from $1.83 /hr”.

Serverless rates, read on 2026-09-09:

Two of these moved. The H100 PCIE rate was $0.000986 per second when this entry was last checked on 2026-09-02 and is now $0.000972, and the on-demand H100 teaser went from $1.74 to $1.83 per hour. The RTX 5090 row is new. Beam’s own worked example puts an RTX 4090 endpoint with 2 physical cores and 16 GiB RAM at $1.77 per hour.

What is gone besides the credit

The $0 base plan and the $89 per month Team plan are both absent from the rebuilt page, and so are the free-plan concurrency caps this entry previously reported (5 GPU containers, 30 CPU containers). Nothing on the current page replaces them with different numbers; the plan structure simply is not published any more. Treat the old caps as unverified rather than as still in force.

What is still free

One thing survives and it is stated plainly in the docs: “Storage is included free up to 1 TB. Beyond that, it costs $0.021 per GB per month.” Storage volumes are not billed as compute, so data sitting in a volume while containers are stopped costs nothing under 1TB. That is the entire free component as of 2026-09-09.

The keep-warm defaults that quietly bill

Unchanged and still worth knowing before the first bill. Containers stay running after each request and that time is charged: endpoints, ASGI and realtime for 180 seconds, task queues for 10 seconds, pods for 600 seconds. The docs are explicit that keep_warm_seconds is billed, and their worked example bills 301.1 seconds for a request that did 100ms of actual work. With no monthly credit absorbing it, a handful of test requests now bills real money from the first one.

What this means if you read the $30 somewhere else

The figure was real. Beam’s pricing page carried “$30 Free Credit, Every Month” and the wording “refreshed monthly”, and this site verified it on 2026-08-21, 2026-08-29 and 2026-09-02. The daily source check found it missing from every cited page on 2026-09-09, and a manual pass over the pricing page, both docs pages, the homepage and the docs llms.txt index found no free credit wording anywhere. What is not knowable from the outside is whether the allowance still exists for accounts that had it, or inside the signup flow: this entry reports what official public pages state, and they now state nothing.

For a serverless GPU platform that does still publish a recurring monthly credit, Modal remains at $30 per month on its $0 Starter plan, with a payment method required on file. Cerebrium is the closest comparison in the other direction: a free plan where every compute second is billed.

The benchmarking ban is gone, with the terms that carried it

Beam replaced its terms of service on 2026-09-14, per the effective date on the document itself; the version it replaced was dated 2023-01-23. The restriction this entry used to quote, “You may not: … (viii) benchmark or perform competitive analysis on the Service.”, is not in the new document, and neither the word benchmark nor the phrase competitive analysis appears anywhere in it.

The replacement is section 3.1, “Restrictions”, an eight-item list running (a) through (h): derivative works, reverse engineering, avoiding fees or exceeding plan quotas, reselling the service as a hosting or proxy offering, circumventing security or metering, probing or scanning for vulnerabilities without written authorization, removing proprietary notices, and degrading the service for other customers. Measuring performance is not among them.

What that changes here: this entry previously declined to publish measured timings for Beam because the old terms forbade it. That reason no longer exists. It has not been replaced by a measurement, because nothing has been tested yet, and the whatYouCanDo field still says so. Item (f) still bars probing or scanning for vulnerabilities without written authorization, which is a different activity from timing a GPU.

One navigation note, unchanged since 2026-08-22: the terms link in the beam.cloud footer points to the docs site, and the path www.beam.cloud/terms-of-service returned a 404 on that date.