← All providers
Groq pricing
verified 2026-09-02Custom LPU hardware serving open models at very high speed. Often the cheapest way to run Llama-class models fast.
data residency & retention
from Groq's own policies · verified 2026-09-02Headquarters
🇺🇸 United States (US)
Where inference runs
US, EU, Other
Groq discloses a global inference fleet — US, Helsinki Finland, Canada, and Dammam Saudi Arabia — but publishes no mapping of which site serves a given API request.
Region pinning
Enterprise plans only
Region-specific endpoints are enterprise-only; free/developer tiers use the global endpoint, and any retained data is stored in US-located GCP buckets regardless of compute location.
Prompt retention
Not retained
By default Groq does not retain customer data for inference requests, with a carve-out to temporarily log inputs/outputs for troubleshooting or abuse investigation, retained up to 30 days.
Trains on API data
No
The services agreement bars Groq from using inputs or outputs for training or fine-tuning any models unless explicitly permitted by the customer.
Zero data retention
Available
Self-serve ZDR toggle in Console Data Controls, which disables even the troubleshooting/abuse logging carve-out.
current models · $ per 1M tokens
| Model | Input /1M | Cached in | Output /1M | Context |
|---|---|---|---|---|
GPT OSS 120B GPT-OSS | $0.15 | — | $0.60 | 131K |
GPT OSS 20B GPT-OSS | $0.07 | — | $0.30 | 131K |
Safety GPT OSS 20B GPT-OSS | $0.07 | — | $0.30 | 131K |
Qwen 3.8-27B Qwen 3.8 | $0.80 | — | $4 | 131K |
Qwen 3.6-27B Qwen 3.6 | $0.60 | — | $3 | 131K |
Llama Prompt Guard 2 86M Llama Guard | $0.04 | — | $0.04 | 1K |
Llama Prompt Guard 2 22M Llama Guard | $0.03 | — | $0.03 | 1K |
Groq totals · 1M in + 1M out
Llama Prompt Guard 2 22M
$0.0600
Llama Prompt Guard 2 86M
$0.0800
GPT OSS 20B
$0.3750
Safety GPT OSS 20B
$0.3750
GPT OSS 120B
$0.7500
Qwen 3.6-27B
$3.60
Qwen 3.8-27B
$4.80
Cost of 1M input + 1M output tokens. Bar length is square-root scaled so cheap models stay visible next to premium ones.