Introducing preemptible compute: the same compute, half the price

Together AI's preemptible GPU service trades uninterrupted access for a lower rate. Its public-preview announcement prices spare-capacity nodes at half the on-demand rate, but allows the company to reclaim them. A workload receives a termination signal and up to five minutes to save progress and exit.

That makes recoverability the deciding factor: a batch job that resumes from saved progress can use the discount, while work that loses everything when interrupted may waste its savings. The preview promises no minimum node lifetime. Replacement capacity is requested automatically, but arrives only as spare resources become available; critical components can remain on standard nodes in the same cluster.