Groq Review: Features, Pricing and Who It Suits Best
A Groq review for late 2026: LPU speed, GPT-OSS pricing, free-tier limits, the smaller model catalog, and who the platform suits best — read the verdict.
By Galactic News Editorial · Published 2026-11-09T10:21:04+00:00 · Updated 2026-11-09T10:21:04+00:00
Groq Review: Features, Pricing and Who It Suits Best
Groq is an inference platform that runs open-weight language models on its own custom hardware, called the LPU. Its pitch is simple: very fast responses at low per-token prices. This Groq review covers what the platform offers, what it costs in late 2026, and who should use it.
What Groq Is
GroqCloud is the hosted API service. You send requests to an OpenAI-compatible endpoint and get fast completions back. The base URL is https://api.groq.com/openai/v1, so existing OpenAI client code works with a small change. Groq also runs Whisper for speech-to-text and offers text-to-speech in preview.
The company behind the models is not always Groq itself. Groq serves open-weight models released by others, such as OpenAI's GPT-OSS family. The hardware is Groq's own. The models are not.
Groq also builds Compound systems, which combine models with tools. For most developers, though, the product is the fast chat API.
Groq Review: Speed and the Model Catalog
Speed is the reason people choose Groq. The public rate card lists GPT-OSS 20B at about 1,000 tokens per second and GPT-OSS 120B at about 500 tokens per second. Those are the speeds Groq lists next to each model. Real speeds vary with load, prompt length, and output length.
The catalog shrank during 2026. As of late October 2026, the self-serve production text models are GPT-OSS 120B, GPT-OSS 20B, and the preview model Qwen3.8-27B. The Llama models that once defined the platform, Llama 3.1 8B and Llama 3.3 70B, shut down for free and Developer tiers in August 2026. They remain available only on Enterprise plans. That change matters. Old tutorials that call a Llama model ID will fail on a current Developer account.
Groq also lists Whisper for transcription and Orpheus text-to-speech in preview. Audio is priced per hour, not per token.
Groq Pricing in 2026
Groq's pricing page now redirects to the homepage. Per-model prices live in the models page of the GroqCloud documentation. The published rates as of late October 2026 are:
| Model | Input per 1M tokens | Output per 1M tokens | Listed speed |
|---|---|---|---|
| GPT-OSS 20B | $0.075 | $0.30 | About 1,000 tokens per second |
| GPT-OSS 120B | $0.15 | $0.60 | About 500 tokens per second |
| Qwen3.8-27B (preview) | $0.80 | $4.00 | About 450 tokens per second |
Two discounts cut those rates. Prompt caching bills repeated prefix tokens at 50 percent off on the GPT-OSS models. The Batch API also takes 50 percent off in exchange for a slower turnaround window. Audio pricing is separate: Whisper v3 costs $0.111 per audio hour, and Whisper Turbo costs $0.04 per audio hour.
There are three tiers: a Free plan, a Developer plan that pays as you go, and Enterprise. There is no seat fee. You pay for what you use.
The Free Plan and Its Limits
The free plan is genuinely usable for testing. On GPT-OSS 120B it allows about 30 requests per minute, 1,000 requests per day, 8,000 tokens per minute, and 200,000 tokens per day. Those are base limits, and Groq notes that exact limits live in the console limits page for each organization.
The free plan is not a production plan. The limits are tight enough that a small app with real users will hit them. Its job is to let you test speed and quality before you add a payment method.
Strengths and Weaknesses
Groq's strengths are real and narrow.
- Speed: the LPU hardware delivers some of the fastest token rates in hosted inference. Interactive apps feel instant.
- Price: $0.075 per million input tokens for the 20B model is cheap by any standard. Caching and batch discounts cut it further.
- Compatibility: the OpenAI-compatible endpoint means most existing code works with a base URL change.
- Simplicity: three tiers, per-token billing, no seat fees, and a free plan for testing.
The weaknesses matter just as much.
- Small catalog: two paid text models is the shortest public rate card among major inference hosts. If you need a specific model, Groq may not have it.
- Catalog churn: the Llama shutdown in August 2026 broke existing integrations. Model availability is not a promise.
- Enterprise wall: several capable models, including the Llama family and MiniMax models, sit behind an Enterprise quote with no published price.
- Best-effort free tier: free plan requests can be throttled under load, so latency is not guaranteed.
Who Groq Suits Best
Groq suits developers building latency-sensitive apps on open-weight models. Chatbots, coding assistants, live summarizers, and voice agents all benefit from the speed. If your app streams tokens to a user who is waiting, Groq is hard to beat on price and speed together.
It suits cost-sensitive teams too. A 20B model at $0.075 per million input tokens handles drafting, classification, and extraction cheaply. Pair it with prompt caching and the bill drops further.
It does not suit teams that need a wide model catalog, a specific frontier model, or a stable long-term model list. It also does not suit teams that need guaranteed latency on a free tier. For those needs, a larger provider with a broader catalog is the safer choice.
Bottom Line
Groq is the fastest cheap way to run open-weight models, and the thinnest catalog among major hosts. GPT-OSS 120B at $0.15 input and $0.60 output per million tokens, with a free tier for testing, makes it easy to try. The trade is choice: two paid text models, with the Llama family behind an Enterprise wall since August 2026. If speed and price matter more than model variety, Groq is a strong pick. If you need a specific model, check the catalog before you build.