Trainable models
Every trainable model can also be evaluated and deployed. Evaluation uses the same rates, without the train charge.Deployments
Deployments are billed per token at the base model’s prefill, cached prefill, and sample rates. There is no hourly charge.Evaluation models
These models can’t be trained. Use them for evaluation and grading. GPT-6 Astra, GPT-5.6 Sol, GPT-5.6 Luna, GPT-5.4 Mini, Claude Opus 5.5, and Claude Sonnet 5.5 are also on the inference API. GPT-5.6 Terra is available on the inference API, but not for evaluation or grading.
Claude cache writes are 5-minute; 1-hour writes cost 2× prefill.
Long-context prompts
Requests whose prompt exceeds the threshold are billed entirely at these rates.Grading
Model-based graders are billed at the grader model’s token rates above.Sandbox environments
Each trajectory runs in its own sandbox, billed per second from ready to stop.
Set
cpus and memory_mb in env_resources. Sandboxes can burst
above their request, so you can size below peak. Each second is billed at the
greater of the request and actual usage.
Example: 1 CPU and 4 GiB cost $0.36/hour, so 1,000 ten-minute trajectories cost $60.