Robotics / VLA

Pricing in this category is not yet standardized. Most of the models on this page run on dedicated GPU infrastructure — Vast.ai, Replicate, self-hosted — and you pay per second of inference compute rather than per call or per token. Plan around €0.001-€0.01 per inference step (one camera frame plus one instruction) on H100-class hardware. A continuous policy running at 10 Hz costs roughly €0.36-€3.60 per hour of robot operation, before energy and supervision costs.

The trade-off triangle is generalization, latency, and physical scope. Larger VLAs (RT-2-X, OpenVLA-7B) generalize to novel objects and instructions but inference at 1-3 Hz, which is too slow for closed-loop dexterous control. Smaller distilled models (Octo, π-0-fast, RDT-1B) hit 30-50 Hz but only generalize within their training distribution. For tabletop manipulation in a controlled cell, the small fast model is usually correct. For research that needs language and visual generalization, the larger model is.

Watch out for the sim-to-real gap: most VLA training data is collected in simulation or on specific robot embodiments. Deploying on a different arm, gripper, or camera geometry typically requires fine-tuning on a few hundred to a few thousand new demonstrations. Also watch out for safety — these models occasionally output unsafe joint trajectories; always run a low-level safety filter (joint limits, force limits, workspace bounds) between the policy and the hardware.

Top picks above cover the most generalizable research flagship, the cheapest run-on-shared-GPU option, the largest open-weights model, and the fastest realtime control policy. Commercial managed-API offerings will be added as providers launch them.