# Groq

> Fast, OpenAI-compatible inference API for open models, running on custom LPU chips

- Category: [Developer Tools](https://finderatlas.com/categories/developer-tools)
- Pricing: Freemium
- Website: https://groq.com
- Platforms: Web, API
- API: available
- Made by: Groq LLC, founded 2016, San Francisco, CA
- Last updated: 2026-10-20
- FinderAtlas page: https://finderatlas.com/tools/groq

## Overview

Groq runs an inference cloud, GroqCloud, that serves open-weight language, speech and safety models through an API that follows the OpenAI format. Requests run on Groq's own Language Processing Unit (LPU) chips instead of GPUs, and the company positions the service around low latency and high token throughput. Its model list includes OpenAI GPT-OSS 120B and 20B, Qwen3.8-27B, Whisper speech-to-text, Orpheus text-to-speech and Llama Prompt Guard.

## Key features

- LPU inference
- OpenAI-compatible API
- Open model catalog
- Speech models
- Built-in tools and MCP
- Batch and Flex tiers
- Structured outputs and reasoning
- Spend limits and projects
- LoRA inference

## Pricing

- **Free** ($0): Free API access with lower rate limits, for evaluation and small projects
- **Developer** ($0.075-$4.00/1M tokens): Pay-as-you-go usage billing with higher rate limits, no monthly fee
- **Enterprise** (Custom): Contact sales for dedicated capacity and enterprise-only models

## Best for

- Low-latency LLM API calls
- Real-time voice and agent apps
- Migrating OpenAI-style code to open models

## Who uses it

- Software developers
- AI application builders
- Enterprise engineering teams
- Voice and agent developers

## Strengths

- Listed speeds of about 280 to 1,000 tokens per second on hosted models
- OpenAI-compatible API and Responses API support, so migration is mostly a base URL change
- Free tier with an API key, and pay-as-you-go pricing with no monthly fee
- Low per-token prices, for example $0.15 input and $0.60 output per 1M tokens for GPT OSS 120B
- Includes speech-to-text and text-to-speech models alongside text models

## Limitations

- Only hosts a curated set of open models; no proprietary frontier models such as GPT or Claude
- Some models, including Llama 3.3 70B, are enterprise-only and require contacting sales
- Free tier rate limits are low for production use
- No public pricing page on groq.com; prices sit in the developer docs
- Preview models can be discontinued at short notice

## Not recommended for

- Training or fine-tuning models
- Users who need proprietary frontier models

## Integrations

- OpenAI SDK
- Factory Droid
- OpenCode
- Kilo Code
- Roo Code
- Cline
- Google Workspace
- MCP connectors

## FinderAtlas editorial scores (out of 10)

- Inference Speed: 9.5/10
- Latency: 9.2/10
- API Compatibility: 9/10
- Cost Efficiency: 8.5/10
- Model Selection: 7/10

## Alternatives

- [Browserbase](https://finderatlas.com/tools/browserbase): Cloud browser infrastructure and APIs that let AI agents browse and act on the web ([compare](https://finderatlas.com/compare/browserbase-vs-groq))
- [Browser Use](https://finderatlas.com/tools/browser-use): Open-source library and cloud platform for AI agents that control real web browsers ([compare](https://finderatlas.com/compare/browser-use-vs-groq))
- [Warp](https://finderatlas.com/tools/warp): Open-source terminal with built-in AI agents and cloud infrastructure for coding workflows ([compare](https://finderatlas.com/compare/groq-vs-warp))
- [Greptile](https://finderatlas.com/tools/greptile): AI code review that checks every pull request against your full codebase context ([compare](https://finderatlas.com/compare/greptile-vs-groq))
- [Langfuse](https://finderatlas.com/tools/langfuse): Open-source platform to trace, evaluate and improve LLM applications and agents ([compare](https://finderatlas.com/compare/groq-vs-langfuse))
- [fal](https://finderatlas.com/tools/fal): Run image, video, audio and 3D generative models through one fast developer API ([compare](https://finderatlas.com/compare/fal-vs-groq))
- [Replicate](https://finderatlas.com/tools/replicate): Run, fine-tune and deploy open-source AI models through a pay-per-use API ([compare](https://finderatlas.com/compare/groq-vs-replicate))
- [OpenRouter](https://finderatlas.com/tools/openrouter): One API and one bill for hundreds of AI models across many providers ([compare](https://finderatlas.com/compare/groq-vs-openrouter))

## Recent updates

- 2026-08-17: Groq announced a $350 million Series A led by Disruptive at a $3.5 billion valuation, with planned participation from...

## FAQ

**What is Groq?**

Groq is an AI inference company. Its GroqCloud service hosts open-weight models behind an OpenAI-compatible API and runs them on the company's own LPU chips.

**Is Groq the same as Grok?**

No. Groq (with a q) is the inference provider described here. Grok is the chatbot from xAI.

**Is there a free plan?**

Yes. You can get an API key and use hosted models on the Free tier, which has lower rate limits than the Developer tier. Upgrading to Developer needs a payment method, and you pay only for tokens used.

**How is Groq billed?**

Usage-based. Text models are priced per million input and output tokens, Whisper per hour of audio and Orpheus per million characters. The bill arrives at month end, or earlier when cumulative usage crosses the $1, $10, $100, $500 and $1,000 thresholds.

**Can I use my existing OpenAI code with Groq?**

Mostly. Groq documents OpenAI compatibility: set the base URL to https://api.groq.com/openai/v1, use a Groq API key and choose a Groq model ID. The Responses API is also supported.

**Can I fine-tune or train models on Groq?**

GroqCloud is an inference service. The docs describe LoRA inference for adapters, but it is not a training platform.

**How did the Nvidia deal affect GroqCloud?**

In December 2025 Groq and Nvidia signed a non-exclusive inference technology licensing agreement, and some Groq leaders joined Nvidia. Groq says it remains independent and that GroqCloud keeps running. Press reports mention some customers seeing a narrower model choice afterwards.