Frontier-model post-training on real coding environments can cost about $50K per run

willccbb · x · 2026-07-29

A reposted talk/article argues that post-training a frontier-sized model on a real coding environment is less out of reach than it sounds.

Key points:

In one GLM-5 step, the team reportedly ran 28 nodes in under five minutes at 131k context.
A 1,000-step run can finish in about three days for roughly $50K in rental costs.
The workshop focuses on the open-source tooling built by Prime Intellect, including verifiers and Prime RL.

The main thesis is that evals are the entry point: an environment and an eval are the same unit of logic, so teams that already build product-hygiene evals are partly set up for post-training. The thread also describes a verifier redesign that splits environments into three composable parts—task set, harness, and runtime—so they can be mixed and matched for different training jobs.

Original post →

Frontier-model post-training on real coding environments can cost about $50K per run

More from coding & agent