What is Groq (GroqCloud)? Free API, rate limits, models, pricing, Whisper, Turkish support and data
On Groq (GroqCloud) you try an open model in the browser through the Playground or call it from code through the API; the model responds very fast. It is an inference service. The Free plan needs no payment method but has per-model request and token limits; the Developer plan is pay-as-you-go with much higher limits. Llama 3.1 8B and Llama 3.3 70B were deprecated on August 16, 2026; gpt-oss and Qwen are recommended instead. There is no Turkish text-to-speech. Customer input is not used for training unless permitted. It is not Elon Musk's Grok.
- What it does
- API that runs open models very fast
- Turkish text-to-speech
- No (Orpheus: English, Arabic)
- Training
- Input/output not used for training
- Free plan
- Yes: no card, per-model limits
- Price
- Developer: pay-as-you-go, per token
What does Groq do?
!Groq (GroqCloud, groq.com) and xAI's Grok chatbot are different things.
Groq runs open models (such as OpenAI gpt-oss, Qwen and Whisper) very fast on its own chips. It offers the Playground for trying them without code and an OpenAI-compatible API for connecting them to an application. Its relationship with NVIDIA has been widely covered in the news: according to official information, NVIDIA did not acquire the company; a non-exclusive licensing agreement was made, and Groq continues to run GroqCloud independently.
Who it helps: developers who need fast responses (voice agents, chat applications); people who don't want to run a model on their own server.
How to get a Groq API key
- Open console.groq.com/login; continue with Google, GitHub, SSO or email (you must be 18 or older; it does not say whether phone or ID verification is required).
- Create an API key on the console.groq.com/keys page. The Free plan needs no payment method.
- Environment variable:
export GROQ_API_KEY=<key> - Library:
pip install groq
It also works with the OpenAI SDK: set base_url to https://api.groq.com/openai/v1. Some fields such as logprobs and logit_bias return a 400 error. A desktop or mobile app is not mentioned in the official documentation.
Which models does Groq have? Was Llama removed?
| Model | What it's for, per Groq | Note |
|---|---|---|
| openai/gpt-oss-120b | Open-weight flagship model; tool use, reasoning, structured output | Production model; ~500 tokens/s |
| openai/gpt-oss-20b | Smaller model; coding and reasoning | Production model; ~1000 tokens/s |
| qwen/qwen3.8-27b | Multimodal (text and image), agentic coding | Preview; don't use in production |
| whisper-large-v3 and -turbo | Speech → text, multilingual | Turbo is transcription only |
| Orpheus (English, Arabic) | Text → speech | Preview; no Turkish |
!llama-3.1-8b-instant and llama-3.3-70b-versatile were deprecated on August 16, 2026 (Free and Developer tiers); gpt-oss-20b, gpt-oss-120b or qwen were recommended instead. Older articles and some examples still mention Llama.
The model list changes often; this section was checked on October 14, 2026. Current prices are on the console.groq.com/docs/models page (groq.com/pricing now redirects to the home page).
Is Groq free? Limits and pricing
- No payment method needed
- Per-model request and token limits
- Playground and API
- Batch, Flex, spending limit
- High limits
- Much higher limits
- Batch (50% discount), Flex (10x limits)
- Spending limit, chat support
- Performance tier, custom contract
| Free | Developer | |
|---|---|---|
| Requests per minute | 30 | 1,000 |
| Requests per day | 1,000 | 500,000 |
| Tokens per minute | 8,000 | 250,000 |
| Tokens per day | 200,000 | No limit |
The numbers come from the HTML tabs of the rate-limits page; the .md version only gives the Free table. Your exact limits are on the Limits page in the console.
Prices are in US dollars, excluding tax, per token (e.g. gpt-oss-120b costs $0.15 input and $0.60 output per million tokens). Payment: card, US bank account, SEPA; you are not charged immediately, an invoice arrives at month-end or when thresholds are reached. The billing FAQ says “unlimited requests”, but the limits page gives numbers for Developer too.
Official pricing page last checked on .
Does Groq work in Turkish?
| Text responsesUnclear | The response language depends on the model you connect; there is no official source on Turkish quality. |
|---|---|
| Speech → text (Whisper)Unclear | It says multilingual; Turkish accuracy is not officially stated. |
| Text → speechNo | Orpheus is English and Arabic (Saudi) only. |
Where is my data, and is it used for training?
Services Agreement 4.2: Groq may not use input and output for model training or fine-tuning unless the customer permits it.
Inference requests are not stored by default; usage metadata (which contains no input or output) is kept. Reliability logs for up to 30 days, Batch files for 30 days. Data is in the US (Google Cloud). Settings: console.groq.com/settings/data-controls.
The agreement says the DPA and BAA do not apply to beta services unless explicitly stated. Whether “preview” and “beta” are the same is not stated; check this before processing personal data.
There is no text specific to KVKK (Turkey's Personal Data Protection Law); there are standard contractual clauses and a DPA. Account deletion steps are not documented.
Common problems and fixes
| What happens | What to do |
|---|---|
| 429 Too Many Requests | Rate limit exceeded: wait for the retry-after period and slow down your requests. For a lasting need, use the Developer plan; see your own limits on the Limits page in the console. |
| After a while, 429 comes back | It may be the Free plan's per-minute or daily limit. Users say the Free limits are low. |
| 400 blocked_api_access | The monthly spending limit was reached; raise or remove the limit. It resets on the 1st of the month. |
| 498 capacity_exceeded (Flex) | Flex capacity is full; wait briefly and retry. |
| An old Llama model doesn't work | It was removed on August 16, 2026; switch to gpt-oss-20b or gpt-oss-120b. |
Frequently asked questions
What is Groq and what does it do?
An inference service (GroqCloud) that runs open models very fast. It offers a Playground in the browser and an OpenAI-compatible API for code.
Is Groq free?
Yes, the Free plan needs no payment method, but there are per-model request and token limits. For example, gpt-oss-120b allows 30 requests per minute and 1,000 requests per day.
How do I get a Groq API key?
Sign in to console.groq.com and create a key on the console.groq.com/keys page.
How much does Groq cost?
The Developer plan is pay-as-you-go, per token; gpt-oss-120b costs $0.15 input and $0.60 output per million tokens. Prices are in US dollars, excluding tax.
Does Groq have Llama?
Llama 3.1 8B and Llama 3.3 70B were deprecated on August 16, 2026. gpt-oss and Qwen models are recommended instead.
Can Groq do Turkish text-to-speech?
No; Orpheus is English and Arabic (Saudi) only. Whisper transcribes audio, but its Turkish accuracy is not stated.
Does Groq use my data for training?
The Services Agreement prohibits using input and output for training unless the customer permits it. Requests are not stored by default.
Are Groq and Grok the same thing?
No. Groq is an inference service; Grok is xAI's chatbot.
Related guides
Sources
Not clear in the official sources: Phone or ID verification at sign-up · account deletion steps · local-currency pricing and student discounts · any country restriction or support statement for Turkey · gpt-oss and Qwen model licenses · a separate Playground limit · Groq Chat's sign-in requirement and limits · whether the DPA applies to preview models. Each is covered in the relevant section along with what you can do.