Problem#
These rate cards disagree with models.dev on input price (measured 2026-10-09). I have not checked which side is right.
| Row | modelschemas (per 1M) | models.dev |
|---|
amazon-bedrock/anthropic.claude-opus-5-5 and global.… | 4.4 | 4 |
amazon-bedrock/eu.amazon.nova-lite-v1:0 | 0.06 | 0.069 |
mistral/mistral-large-4 | 1.36 | 0.68 |
mistral/magistral-medium-latest | 1.5 | 2 |
mistral/ministral-8b-latest | 0.15 | 0.1 |
huggingface/meta-llama/Llama-3.3-70B-Instruct | 1.026 | 0.59 |
huggingface/XiaomiMiMo/MiMo-V2.5-Pro | 0.522 | 1 |
together/Qwen/Qwen3.7-Max | 2.5 | 1.25 |
opencode/gpt-5.6-terra | 2 | 2.5 |
Bedrock has 37 differences in total, Hugging Face 26, Mistral 4.
Ask#
Check each row against the provider's price page. Possible causes to rule out:
- Bedrock: the global profile versus the regional price (regional is often 10% more). The two need separate cards.
- Hugging Face router: the price depends on the upstream host. Say which host the card is for.
- Mistral: a
-latest alias that points at a newer model than the card.
Acceptance#
- Each row is confirmed or fixed, with the price page URL in the PR.