June 13, 2026

Major features

Price cap for Serving

Set a maximum price per GPU for a Serving deployment. Replicas above the cap are flagged and migrated automatically, with an alert if the whole deployment is blocked by the cap.

Bring your own storage bucket

Route inference outputs (sync and async) to your own S3-compatible storage bucket instead of the default. The artifacts gallery also gained delete, a bucket filter, an “scheduled for deletion” indicator, and safer downloads.

External capacity, shown transparently

Serving deployments can burst to external GPU capacity when needed, with the GPU type and price shown transparently on the card — no hidden surcharges.

Policy changes

• The $3 sign-up bonus is back.
• Referral rewards changed from a flat $10 to 15% of your referred friend’s first top-up, up to $25.

Improvements

• Deployment progress now shows explicit stages (pulling image → downloading model → waiting to be ready) with a specific reason if something fails.
• API keys can now be renamed, get an auto-generated default name, and sort newest-first; temporary Playground session keys are hidden from the list.
• Ongoing security hardening across authentication and internal service communication.

Fixes

• Fixed the Quick Deploy page showing an incorrect cost for some CPU/RAM configurations.
• Fixed a mislabeled CPU generation shown in the GPU listing.
• Fixed several Playground crashes (including image-to-video) when a model’s requirements couldn’t be determined or a request errored.
• Fixed result downloads on the Requests page to use a reliable link instead of opening inline in the browser.
• Fixed multi-GPU external rentals occasionally providing fewer GPUs than requested.
• Fixed external capacity not always being released promptly after a deployment was deleted or ran out of use.