Open weight models give you three ways to pay: a hosted API that serves them per token, a rented GPU you run yourself, or hardware you own. This post compares the first against frontier APIs using published prices, and sets up the second in a companion guide. It compares price only. It does not claim any model matches another in quality, which you must test on your own work.
Hosted open weight prices
Together AI's pricing page, fetched 18 October 2026, lists these serverless prices per million tokens. The page does not date itself, and it says displayed prices refer to the lowest resolution settings and actual prices might vary.
| Model | Input | Output |
|---|---|---|
| gpt-oss-120B | $0.15 | $0.60 |
| Qwen3 235B A22B Instruct 2507 FP8 | $0.20 | $0.60 |
| Gemma 4 31B | $0.39 | $0.97 |
| Llama 3.3 70B | $1.04 | $1.04 |
Check each model's licence yourself. The page does not label which models are open weight.
Frontier API list prices for comparison
| Model | Input | Output |
|---|---|---|
| Claude Haiku 5.5 (prompts up to 100,000 tokens) | $0.10 | $0.50 |
| OpenAI gpt-6-luna | $0.10 | $0.50 |
| OpenAI gpt-5.4-mini | $0.75 | $4.50 |
| Claude Sonnet 5.5 | $2 | $10 |
| OpenAI gpt-6-sol | $2 | $10 |
Sources are Anthropic's and OpenAI's pricing pages, fetched the same day.
One workload, priced every way
Monthly workload: 100 million input tokens and 10 million output tokens, no caching, no batch.
| Model | Monthly cost |
|---|---|
| Claude Sonnet 5.5 | $300 |
| OpenAI gpt-5.4-mini | $120 |
| Llama 3.3 70B (Together) | $114 |
| Gemma 4 31B (Together) | $49 |
| Qwen3 235B A22B (Together) | $26 |
| gpt-oss-120B (Together) | $21 |
| Claude Haiku 5.5 | $15 |
Computed as input millions x input price plus output millions x output price. Look at the table twice. The cheapest frontier small tier is in the same range as hosted open weight models, and a hosted open weight model is not automatically cheaper than a small frontier model. Llama 3.3 70B costs more than Haiku 5.5 here.
Where the real differences are
- Caching. Together lists cached input prices for some models (for example $0.35 against $0.60 for Qwen3.5-397B-A17B) but not all, while the frontier vendors publish cache multipliers for every listed model.
- Control. A model you host stays on your hardware, and data does not leave. Safeguard supports bringing your own model, including on-prem and local ones.
- Operations. A hosted per-token price includes the GPUs, scaling and upgrades. Self hosting does not.
- Quality per task. A cheaper model that needs two attempts is not cheaper.
For the third option, see our self hosting cost per million tokens estimate.
Quick check
Does a lower per token price mean a lower bill?
Not always. Look at the model's quality on your task, retries, caching and the ratio of input to output tokens.
Sources
- Together AI pricing, fetched 18 October 2026
- Claude API pricing, fetched 18 October 2026
- OpenAI API pricing, fetched 18 October 2026
