Ultra-low-latency LLM and speech inference API running on custom LPU hardware, built for real-time AI apps. Read our hands-on Groq review and compare the b

Overview

Groq is an inference platform built around its own Language Processing Unit (LPU), custom silicon designed specifically for transformer workloads rather than general-purpose GPUs. The result is extremely high tokens-per-second throughput — many models stream at 500 to 1,000+ tokens per second — which makes it a favorite for latency-sensitive products like live chat, voice agents, and real-time retrieval. You reach it through a standard OpenAI-compatible API at https://groq.com, so most existing SDKs and frameworks drop in with a single base-URL change. The catalog spans open-weight text models (Llama, Qwen, GPT-OSS, Kimi), Whisper speech recognition, text-to-speech, and a Compound system that wires in web search and code execution. In our evaluation Groq’s standout value is predictable, linear, per-token pricing with no seat tax: you pay only for what you generate, which keeps small projects nearly free while scaling cleanly. It is less suited to teams that need private fine-tunes or guaranteed reserved capacity, where a dedicated GPU cloud or self-hosted stack is the better fit. For the common case of ‘I need a fast model endpoint now,’ Groq removes almost all friction.

Key Features

  • LPU-powered low latency — custom inference hardware delivers 500–1,000+ tokens/sec on many models, so streaming responses feel near-instant for chat and voice.
  • OpenAI-compatible API — swap your base URL and keep existing SDKs, LangChain, or agent frameworks working unchanged.
  • Open-weight model catalog — Llama, Qwen, GPT-OSS, and Kimi text models plus Whisper transcription and Orpheus text-to-speech.
  • Built-in tools — compound systems add web search, website visits, and Python code execution without extra services.
  • Transparent per-token billing — no monthly minimum or seat fee; cached input and Batch API cuts cost roughly in half.

Pricing

PlanPriceFor
Free / Developer$0Pay-as-you-go, generous free rate limits on small models
Production (per token)$0.05–$1.00 / 1M in, $0.08–$3.00 / 1M out by modelProduction apps billed purely on usage
AudioWhisper Turbo $0.04/hr, V3 $0.111/hrSpeech transcription workloads
EnterpriseCustomReserved capacity, support, compliance

Comparison

Compared to Exa, Groq is an inference engine rather than a search index — Exa returns retrieved documents, while Groq runs the model that reasons over them, and the two are frequently paired in the same RAG stack. Against Perplexity, Groq is the raw model layer underneath: Perplexity ships a finished cited-answer product, whereas Groq gives you the speed to build your own answering experience.

Compare alternatives

Side-by-side with the 2 closest alternatives.

ToolCategoryPricingVisit
Groq (this) code, searchFree $0 · From $0.04/mo Site ↗
ExasearchFree $0 Site ↗
PerplexitysearchFree $0 · From $20/mo Site ↗
Groq Current

Ultra-low-latency LLM and speech inference API running on custom LPU hardware, built for real-time AI apps. Read our hands-on Groq review and compare the b

codesearch
Free $0 · From $0.04/mo

Neural search API that finds pages by meaning for AI apps. Neural / embeddings search Read our hands-on review and compare the top AI Search alternatives.

search
Free $0

AI-native answer engine that responds with cited, real-time sources instead of a list of blue links. Cited, source-backed answers Read our hands-on review

search
Free $0 · From $20/mo
Editor’s Review
4.4/5
Pros
  • +Tokens-per-second speeds up to 1,000+ make streaming feel instant
  • +Pay-as-you-go per-token pricing with no monthly minimum
  • +Open-weight models (Llama, Qwen, GPT-OSS) and Whisper transcription
Cons
  • Throughput-focused hardware trades some peak model quality for speed
  • No long-term fine-tuning or dedicated private deployments on the free tier

Groq is the go-to choice when latency matters more than squeezing out the last point of benchmark quality. For chat assistants, voice agents, and live RAG it delivers a responsiveness that GPU cloud APIs struggle to match. The catch is that you are renting speed, not control — serious customization still points back to self-hosting.

See all reviews →