Which channel you buy through changes the price, the version you get, who holds your data, and whether legal lets you ship. Price is the smallest of those differences.
These are all real, current, and correct — for the channel they were quoted on. That is exactly why a price without a channel attached is not a price.
| Model | Channel | Input /M | Cached /M | Output /M | Checked |
|---|---|---|---|---|---|
| DeepSeek V4 Profirst party | platform.deepseek.com | $1.32 | $0.044 | $3.96 | 2026-09-26 |
| DeepSeek V4 Prosame model, off-peak | platform.deepseek.com | $0.66 | $0.022 | $1.98 | 2026-09-26 |
| DeepSeek V4 Pro 0813inference platform | Together AI | $1.32 | $0.13 | $3.96 | 2026-09-26 |
| GLM-5.3hosted by another lab | Mistral AI (EU) | $1.40 | unknown | $4.40 | 2026-09-26 |
| GLM-5.3inference platform | Together AI | $1.40 | $0.26 | $4.40 | 2026-09-26 |
| gpt-oss-120Bno first-party endpoint | Together AI | $0.15 | unknown | $0.60 | 2026-09-26 |
Off-peak for DeepSeek is half price and applies outside 01:00-04:00 and 06:00-10:00 UTC on weekdays. Same model, half the money, decided by clock time.
Ask these before the security review, not during it. Each one has killed a deal we have watched.
Promotional rates that end. GPT-5.6 Sol is currently $4 / $20 against a list price of $5 / $30, and OpenAI says the promotion holds at least through 21 November 2026. Model the list price. A rate with an expiry date is a decision you have to revisit, whether or not you planned to.
Version lag on marketplaces. A hyperscaler SKU with the same model name is not guaranteed to be the same build as the lab serves. Confirm the version string before you benchmark, or you will spend a week explaining a quality regression that is really a version difference.
Cached-input asymmetry. Two hosts can list identical input and output rates and differ several-fold on cache hits. If your workload has a stable prefix, the cached column is the one that decides the invoice.
Regional multipliers. Data residency is not free anywhere: OpenAI charges +10% for regional processing, Mistral +10% for regional inference, and Anthropic bills US-only inference at 1.1x. Budget for it rather than discovering it on the first invoice.
Licence drift on "open" models. Apache 2.0 and MIT are permissive. Modified MIT, bespoke community licences, and "weights available" with conditions are not. Read the LICENCE file in the model repository — several models in our library carry licences we have not yet verified, and we say so on each record.
The thing nobody budgets: migration. Every lab deprecates models. Ask how much notice you get and what the migration path is, because you will use it.