AI API Models Prices
Last updated: September, 2026
1. What will you use it for?
2. Where will it run?
3. What matters more?
How to choose an open-weight model
1. Start hosted, decide later
The easiest path is to use an open model through the API first. If it works well and you later need full control, you can run the same model yourself without changing your code.
2. Match the size to your hardware
Bigger models are smarter but need much more graphics memory. Every model page lists how much memory each version needs, so you can check it fits your machines before you commit.
3. Check the licence
Open weights don't always mean free for every use. Read the licence on the model page if you plan to use it commercially.
Frequently asked questions
What does open-weight mean?
The model's weights are published, so anyone can download and run it on their own hardware.
Is it cheaper to host it myself?
Only at high, steady volume or when GPUs are already idle. For most teams the hosted API is cheaper and simpler.
Will my code change if I move to self-hosting?
No, if your server exposes an OpenAI-compatible endpoint (vLLM and SGLang both do).
What does the rating for open-weight models mean?
Each model is rated from 1.0 to 5.0 specifically for open-weight. Green is 4.0 and above, yellow 3.0–3.9, red below 3.0. The same model can score differently in another category.
How is the rating calculated?
It comes from public data, not paid reviews: 50% production adoption (public leaderboard standing, routing popularity and download volume), 30% the benchmark that matters for this category, and 20% tooling and value for money (tool calling, structured outputs, context size and price against the category median).
What does the ✓ next to a rating mean?
It means an independent lab — LMArena or Artificial Analysis — has tested the model. Ratings without it are estimated from published specs such as context size, tooling support and price.
How are brand-new frontier models rated?
When a new model has strong independent benchmark results but no public usage data yet, adoption is left out instead of counted as low, and the rating uses 60% benchmark and 40% tooling and value.
How often are ratings updated?
Every week. If a source can't be reached, the previous week's figure is kept.
