Groq is a fast inference platform for open AI models, built on an in-house chip called the LPU and reachable through the GroqCloud API. Over 5 million developers run Llama, Qwen, GPT-OSS or Whisper on it, often beyond 500 tokens per second. The entry tier costs nothing and asks for no credit card; production traffic is billed per million tokens, with no subscription.
The LPU (Language Processing Unit) is an application-specific integrated circuit with a single job: serving language models with minimal latency and deterministic timing. Jonathan Ross, who worked on Google's first TPUs, founded the company in 2016 on a bet that GPUs, inherited from graphics rendering, would never be the right machine for inference.
Throughput ranges from 250 to over 1,000 tokens per second depending on the model, with energy efficiency the company rates at up to ten times that of GPUs. The bet paid off: in December 2025, Nvidia signed a non-exclusive license worth roughly $20 billion for the architecture, and Groq 3 LPX racks packing 256 chips are planned alongside its Vera Rubin platform.
GroqCloud serves open-weight models only: Llama, Qwen, GPT-OSS, plus Whisper for transcription and synthetic voices. Nothing is trained in-house; Groq acts as an execution engine for other labs' work, and Meta handed it part of the inference behind the official Llama API in 2025.
Switching is a formality: the API mirrors OpenAI's, so you swap the base URL and the model name, and your app starts answering from LPUs (by the time the server restarts, the first reply is already back). Plenty of headroom to power coding assistants or a voice agent without touching the rest of your stack.
An API key takes two minutes to create on console.groq.com, no card required, with rate limits set per model. Registering a payment method lifts those ceilings and moves billing to pure consumption. The Batch API halves every rate, prompt caching halves repeated input costs, and both discounts stack.
Models rotate in and out of the catalog at a steady pace, so check the pricing page on groq.com before budgeting anything; it is the only source that counts.
| Model served | Input ($/M tokens) | Output ($/M tokens) |
|---|---|---|
| GPT-OSS 20B | $0.075 | $0.30 |
| GPT-OSS 120B | $0.15 | $0.60 |
| Qwen 3.6 27B | $0.60 | $3.00 |
Yes, GroqCloud includes a free tier with no credit card: every model in the catalog is available, capped by per-minute and per-day rate limits. Adding a payment method then moves the account to pay-as-you-go billing per million tokens, with no subscription and no monthly minimum.
Nothing connects the two companies: Groq, founded in 2016, designs chips and sells inference, while Grok is the chatbot from xAI, launched by Elon Musk in 2023. Groq holds the trademark and publicly asked xAI to pick another name when the confusion began.
No, the catalog sticks to open-weight models such as Llama, Qwen and GPT-OSS. Closed models stay with their publishers, which explains a common setup: Groq handles the speed-sensitive layer of the app, and a second provider covers the heaviest reasoning.
Day to day, nothing changes: the December 2025 agreement is a non-exclusive license on the LPU architecture valued around $20 billion, and GroqCloud keeps operating independently under CEO Adam Winter, who took the role in 2026. Founder Jonathan Ross moved to Nvidia, while the platform, its API and its pricing stayed in place.
Verdict: Up to nineteen times cheaper than a comparable OpenAI tier, with replies that land before a loading spinner would: keep a closed model on the bench for the very hardest reasoning, and Groq becomes the natural home for developers shipping voice agents, live chatbots and heavy transcription pipelines.
Claude Fable 5 
Grok 4.6 
Claude Opus 4.8 
