Unleash the Power of LLM Inference with Groq API

Groq runs open-weight models on its own LPU inference chips and sells access per token through GroqCloud. The API is OpenAI-compatible: point an OpenAI SDK at https://api.groq.com/openai/v1, swap in a Groq key, and existing chat, Responses and audio code works. Groq's docs list text generation, reasoning, tool use, structured outputs, OCR and image input, speech to text (Whisper), text to speech (Orpheus), content moderation and prompt caching, plus built-in code execution and browser search tools, remote tools over MCP, and connectors such as Google Workspace.

The self-serve catalog is short and changes often. On November 1, 2026 the production list showed GPT-OSS 120B and 20B plus Whisper Large v3 and v3 Turbo; preview models included Qwen3.8 27B, GPT-OSS-Safeguard 20B, Llama Prompt Guard and Orpheus TTS. Llama 3.1 8B Instant and Llama 3.3 70B Versatile are still listed but marked Contact Sales. You sign up in the console, start on the Free plan and move to the pay-as-you-go Developer plan by adding a payment method.

Groq no longer has a standalone pricing page (groq.com/pricing redirects to the homepage); per-model prices are published in the models docs.

Worked example: 1M input plus 1M output tokens on GPT-OSS 120B costs $0.15 + $0.60 = $0.75, or about $0.38 through the Batch API, which Groq prices at 50% of synchronous rates (24-hour to 7-day window). Cached input tokens are 50% off on the GPT-OSS models, applied automatically. One hour of audio through Whisper v3 Turbo is $0.04, so 100 hours is $4. Developer billing: new accounts are invoiced as cumulative usage crosses $1, $10, $100, $500 and $1,000, then monthly; nothing is billed under $0.50; cards, US bank accounts and SEPA debit are accepted; refunds are case by case.

Prices checked on the vendor's pricing page on November 1, 2026. See Groq models and pricing.

What Groq lists in its GroqCloud docs:

Groq suits developers who need low-latency responses from open models, such as voice agents, chat UIs and agent loops where seconds matter, and teams transcribing audio in volume, where Whisper Turbo costs cents per hour. The GPT-OSS models at $0.075 to $0.60 per million tokens make it a cheap default for high-volume, moderate-difficulty tasks.

Look elsewhere if you need closed frontier models (GPT, Claude, Gemini) or a wide open-model catalog: OpenRouter routes one key across hundreds of models, and DeepInfra or Together AI host many more open models. If you built on Llama 3.x through Groq's self-serve tier, note those models now require a sales conversation.

Independent write-ups agree on speed and price and focus their criticism on limits and churn. AI Tools Atlas rates Groq a solid choice for fast open-model inference and Whisper pricing, but says free-tier limits push real usage to pay-as-you-go and that deprecations can force mid-project model swaps. The pricing tracker UsagePricing recorded Groq moving Llama 3.1 8B and Llama 3.3 70B from self-serve prices to Contact Sales on August 26, 2026, and removing the Compound agent systems from the catalog by October 24, 2026. The practical takeaway from these sources: Groq is fast and inexpensive for what it hosts, but check the current models page before you depend on a specific model.