Castform Pricing 2026

What Castform charges for

Castform is a reinforcement-learning platform for post-training open models. Its public rate card separates hosted inference from training. The inference row above comes from the structured dataset and retains its own verification date; this page’s original review date does not renew it. The separate training record names Qwen/Qwen3.5-4B, with a 262,144-token context limit, text modality and self-serve availability. Introductory pricing is conditional; the API does not specify an expiry date. The current public pricing page says training meters rollout generation plus weight updates. Base-model inference and serving trained checkpoints use the same inference price. Castform advertises $50.00 in free credits to start your first training run; credit expiry and additional account-specific conditions are not stated or independently verified. Similarly named training and inference endpoints are not interchangeable.

What the Neon benchmark changes

In the dated Neon case study, Castform reports a tuned Qwen3.5-4B retrieval agent outperforming its GPT-5.6 Sol baseline on reward while using a lower rollout budget. Those historical workload costs are not current API prices; see the linked launch analysis for the reported measurements and assumptions.

That is a vendor-reported workload result, not a general model ranking or a guaranteed discount. The public hosted rate does not identify the benchmark checkpoint, and total cost also includes training tokens, synthetic data, embeddings, Neon compute, evaluation, monitoring, and fallback calls. Read the full Castform versus GPT-5.6 Sol analysis.

Labs availability

The tuned retrieval specialist has an explicit Labs blocker. A valid test needs the exact callable artifact, fixed corpus and holdout, published reward weights and stopping rules, plus measured token and database usage. Running the public base endpoint would not test the post-trained result.

Buyer checklist

• Confirm the exact training model and any introductory-rate conditions; do not reuse the inference model’s price.
• Measure cost per accepted answer, including retrieval and fallback infrastructure.
• Keep a blind holdout separate from synthetic training questions.
• Verify the deployed endpoint is the tuned artifact you evaluated.