Groq Review 2026

What is Groq?

GPT OSS 20B on GroqCloud is listed at 1,000 tokens per second, and it costs $0.075 for every million input tokens. That pairing of speed and a small bill is why Groq keeps turning up in developer chats, and it is what this Groq API pricing and speed review checks against the official docs as of October 2026. The company behind it has also changed shape over the past ten months, so the business story matters almost as much as the rate card.

Groq, not Grok: two different companies

The names get mixed up all the time. Groq (with a q) is an American chip and cloud company founded in 2016 and based in Mountain View, California. It sells API access to open-weight models running on its own hardware. Grok (with a k) is the chatbot and model family from xAI, which now brands itself SpaceXSI and sits inside SpaceX. xAI's grok-4.7 model costs $2 input and $6 output per million tokens on the xAI API. Groq does not train or sell Grok, and a Groq key will not call it.

What happened between Groq and NVIDIA

On 24 December 2025 Groq announced a non-exclusive licence of its inference technology to NVIDIA. Founder Jonathan Ross and president Sunny Madra moved to NVIDIA with other team members, and press reports put the deal at about $20 billion, a figure neither company disclosed. It was a licence rather than a purchase. Groq said GroqCloud would keep running without interruption and named Simon Edwards as its new chief executive.

Groq then reinvented itself as an inference cloud. It raised $650 million in June 2026 and, on 17 August, a $350 million Series A at a $3.5 billion valuation led by Disruptive, with NVIDIA planning to take part. It now calls itself an NVIDIA Cloud Partner, and NVIDIA has announced Groq 3 LPX racks built around the LPU design. For an API buyer, the practical point is simple: the service is alive and funded, but the leadership that built the chip now works at NVIDIA.

Groq API pricing per million tokens, model by model

The official prices now live on the models page of the docs, because groq.com/pricing only brings up the homepage. Groq lists them only in US dollars. Rupee figures below are approx., converted at about ₹94.7 to the dollar.

Two things stand out. First, the Llama models that made Groq famous are still served but no longer carry a public price, so self-serve users are steered to gpt oss. Second, preview models such as Qwen 3.8 27B, MiniMax M2.7 and Meta's Muse Spark Open can be pulled at short notice, according to the docs, so do not build a product around them. There are also Orpheus text-to-speech voices at $22 per million characters for English and $40 for Saudi Arabic.

Two discounts apply. The Batch API charges half the synchronous rate for jobs you can wait 24 hours to 7 days for, and prompt caching takes 50% off cached input tokens on the three GPT OSS models, with no code changes. The two do not stack.

Free, Developer and Enterprise in plain terms

Groq runs three account levels. The Free plan needs no payment and gives you every public model at low ceilings. The Developer plan is pay as you go at the rates above, raises limits and opens Batch and Flex processing. Enterprise is a sales conversation, and that is the only route to Llama pricing, MiniMax M2.7 and custom limits.

Flex is the odd one. It costs the same as on-demand but allows ten times the rate limits while spare capacity exists. When capacity runs out, requests fail fast with HTTP 498 and a capacity_exceeded message, so you need retry logic.

Groq free tier rate limits, and why your app hits a 429

The rate-limits page prints one table under a heading that names both the Free and Developer plans, and the text above it calls the figures the Developer plan's base limits. For gpt-oss-120b, gpt-oss-20b and Qwen 3.8 27B they read 30 requests per minute, 1,000 requests per day, 8,000 tokens per minute and 200,000 tokens per day. Whisper gets 20 requests a minute, 2,000 a day and 7,200 audio seconds per hour. Groq does not publish a separate free-plan table, so open the Limits page in your own console to see what your organisation actually has.

The 8,000 TPM figure is the one that bites. A single request with a 6,000-token document plus a long answer can use most of a minute's budget, so a second call fails even though you sent only two. Limits are counted per organisation, not per user, so three teammates testing on one org share the pot. Every response carries x-ratelimit headers for remaining requests and tokens, and a 429 also includes retry-after, which tells you how many seconds to wait. Cached tokens do not count against the limits, and batch jobs have a separate quota. If 429s are constant, ask for higher limits or move work to Batch rather than writing cleverer retries.

Groq vs OpenAI API speed and what it costs you

Groq publishes a throughput figure for each model, which OpenAI does not. OpenAI instead sells speed as a surcharge: Fast mode doubles the token price and Ultrafast on GPT-6.1 Sol and GPT-6 Astra costs six times standard. On Groq, low latency is the default and costs nothing extra.

Price is closer than many expect. Take 20,000 support chats a month at 3,000 input and 500 output tokens each: 60 million in, 10 million out. On gpt-oss-120b that comes to $15 (approx. ₹1,420). On gpt-oss-20b it is $7.50. OpenAI's GPT-6 Luna, at $0.10 input and $0.50 output, lands at $11. So Groq's larger open model is slightly dearer than Luna and the smaller one is cheaper. For speech, Groq's Whisper Turbo at $0.04 an hour is far below OpenAI's gpt-transcribe at $0.0045 a minute (about $0.27 an hour).

OpenAI wins on breadth: GPT-6.1 Sol and Astra, realtime voice, image generation, embeddings and hosted tools on one bill, plus 1.05 million tokens of context against Groq's 131,072. Groq wins when the model is good enough and response time is the product.

Groq LPU vs GPU: where the speed comes from

Groq's Language Processing Unit keeps model weights in on-chip SRAM instead of fetching them from external memory, which is the usual bottleneck when a GPU generates tokens one after another. Groq's platform page says each NVIDIA Groq 3 LPX rack, once released, will link 256 LPU accelerators with 128 GB of on-chip SRAM and 40 PB/s of SRAM bandwidth. The trade-off is capacity: SRAM is small, so large models need many chips, which helps explain why Groq offers a short list of models rather than everything on Hugging Face. LPU inference is good at fast, steady output for chat and agents. GPUs remain the general tool for training and for very large or unusual models.

Prompts, logs and data centres

Groq's Services Agreement, last modified 22 June 2026, says Groq may not use your inputs or outputs to train any model unless you allow it, and that outputs belong to you as customer data. It also bars using the service to build a competing cloud product. By default inference requests are not stored, although inputs and outputs may sit in reliability and abuse logs for up to 30 days. Any customer can switch on zero data retention in Data Controls.

Where is data processed? Groq lists 13 data centres in the US, Canada, Finland, the UK, Saudi Arabia and Australia, and the data page says retained customer data sits in Google Cloud buckets in the United States. There is no India region, which matters for teams under strict localisation rules. Customers in the EEA and Switzerland contract with Groq UK Limited and Saudi customers with Groq Limited; everyone else, India included, deals with Groq LLC in California.

Calling Groq with code you already have

Groq offers an OpenAI compatible API at https://api.groq.com/openai/v1, and the Responses API is supported too. Most apps switch by changing the base URL, the key and the model name. A few OpenAI fields return a 400 error, namely logprobs, top_logprobs, logit_bias and messages[].name, and n must be 1.

• Sign in to the GroqCloud console and create an API key.
• Store it as GROQ_API_KEY on your server, never in a browser or mobile build.
• Point the OpenAI SDK at the Groq base URL and start with openai/gpt-oss-20b.
• Log the rate-limit headers from day one so you know when the free plan stops being enough.
• Move overnight jobs, such as transcribing a backlog with Whisper, to the batch API at half price.

Weak spots before you commit

• A short model list, and 131K context on every self-serve model. No image generation or embedding models appear in the current catalogue.
• Public prices exist only for gpt oss, Qwen preview, Whisper, Orpheus and Prompt Guard. Llama models need a sales call.
• The published 8,000 TPM limit on GPT OSS is tight for document-heavy work.
• USD-only pricing and no India data region.
• The founder and president left for NVIDIA, and the model line-up has changed often, so plan for migrations.

Groq API pricing and speed review: who should pay

Groq suits developers building voice agents, chat widgets, coding helpers and transcription pipelines where a second of delay hurts the experience, and Indian startups that want cheap Whisper transcription for call recordings or regional-language audio. It also makes a handy fallback provider behind an existing OpenAI integration. Teams that need frontier reasoning, long context, images or data stored in India should look at OpenAI's GPT-6 Luna, Anthropic's Haiku 5.5 at the same $0.10 and $0.50 rates, the Google Gemini API, xAI's grok-4.7, or DeepSeek's deepseek-flash at $0.15 and $0.60 off-peak. Start on the free plan and move to Developer only when your logs show you need it.