Groq pricing is per token for chat models and per audio hour for Whisper. gpt-oss-120b costs $0.15 per million input tokens and $0.60 per million output tokens. gpt-oss-20b costs $0.075 and $0.30. The preview model Qwen3.8-27B costs $0.80 and $4.00. Llama 3.1 8B, Llama 3.3 70B, and MiniMax M2.7 are Enterprise models with contact-sales pricing. The Batch API takes 50% off in exchange for a 24-hour to 7-day window, and prompt caching bills repeated prefix tokens at 50% off automatically. groq.com/pricing now redirects to the homepage, so every number on this page comes from the GroqCloud models docs as of October 2026.
Groq API Pricing per Model (October 2026)
These are the production models on GroqCloud. Speeds are the tokens-per-second figures Groq lists next to each model. Rate limits are the Developer plan defaults.
| Model (API ID) | Input $/1M | Output $/1M | Listed speed | Context / max output | Developer limits |
|---|---|---|---|---|---|
| GPT OSS 120B (openai/gpt-oss-120b) | $0.15 | $0.60 | 500 tok/s | 131,072 / 65,536 | 250K TPM, 1K RPM |
| GPT OSS 20B (openai/gpt-oss-20b) | $0.075 | $0.30 | 1,000 tok/s | 131,072 / 65,536 | 250K TPM, 1K RPM |
| Llama 3.1 8B (llama-3.1-8b-instant) | Contact sales | Contact sales | 560 tok/s | 131,072 / 131,072 | Enterprise |
| Llama 3.3 70B (llama-3.3-70b-versatile) | Contact sales | Contact sales | 280 tok/s | 131,072 / 32,768 | Enterprise |
| Whisper Large V3 (whisper-large-v3) | $0.111 per audio hour | n/a | n/a | 100 MB max file | 200K ASH, 300 RPM |
| Whisper Large V3 Turbo (whisper-large-v3-turbo) | $0.04 per audio hour | n/a | n/a | 100 MB max file | 400K ASH, 400 RPM |
TPM is tokens per minute, RPM is requests per minute, and ASH is audio seconds per hour. The two Llama models are still served, but Groq now lists them as Enterprise models with pricing through sales. Some of the price trackers that rank for this query still list retired models such as Gemma 7B, so check the API ID before you budget.
Preview, Safety, and Speech Model Pricing
Groq marks preview models as evaluation-only. They can be discontinued at short notice, so they belong in prototypes rather than production paths.
| Model (API ID) | Price | Listed speed | Context / max output |
|---|---|---|---|
| Qwen3.8-27B (qwen/qwen3.8-27b) | $0.80 in / $4.00 out per 1M tokens | 450 tok/s | 131,072 / 16,384 |
| Safety GPT OSS 20B (openai/gpt-oss-safeguard-20b) | $0.075 in / $0.30 out per 1M tokens | 1,000 tok/s | 131,072 / 65,536 |
| Llama Prompt Guard 2 22M (meta-llama/llama-prompt-guard-2-22m) | $0.03 in / $0.03 out per 1M tokens | n/a | 512 / 512 |
| Llama Prompt Guard 2 86M (meta-llama/llama-prompt-guard-2-86m) | $0.04 in / $0.04 out per 1M tokens | n/a | 512 / 512 |
| MiniMax M2.7 (minimaxai/minimax-m2.7) | Contact sales (Enterprise) | 260 tok/s | 196,608 / 131,072 |
| Orpheus V1 English (canopylabs/orpheus-v1-english) | $22.00 per 1M characters | n/a | 4,000 / 50,000 |
| Orpheus Arabic Saudi (canopylabs/orpheus-arabic-saudi) | $40.00 per 1M characters | n/a | 4,000 / 50,000 |
Batch, Flex, and Prompt Caching Discounts
Batch API: 50% off
The Batch API costs 50% less than synchronous calls. You upload a file of requests and pick a completion window between 24 hours and 7 days. Groq recommends the longer window so jobs finish instead of expiring. Batch requests do not touch your standard rate limits. It covers chat completions with text or vision input, plus audio transcription and translation.
Prompt caching: 50% off cached tokens
Prompt caching runs on supported models with no code change and no fee. When a request starts with the same prefix as a recent one, Groq reuses the computation and bills the cached portion at half price. Cached tokens also do not count toward rate limits. Cache hits are not guaranteed, and cached data lives in volatile memory until it expires. Put the system prompt and tool definitions first and keep them byte-identical across calls to get the most hits.
Flex processing: higher limits at the same price
Flex processing is a service tier for high-throughput work that can tolerate occasional failed requests. It raises rate limits and keeps the same per-token price. Build retries into any client that uses it.
- Evals and nightly jobs: Batch. Half price, and the jobs run outside your live rate limits.
- Agent loops with a long fixed prompt: prompt caching. The repeated system prompt and tool schema bill at half price on every hit.
- Bursty pipelines that can retry: Flex. Same price, more headroom.
Groq Rate Limits and the Developer Plan
The limits in the tables above are the Developer plan base limits from the GroqCloud rate-limit docs. The exact limits for your organization show on the limits page in account settings. Upgrading to the Developer plan raises limits and unlocks Batch and Flex processing. Groq offers higher limits for select workloads.
For an agent, the per-minute token cap decides more than the per-token price. At 250K tokens per minute on gpt-oss-120b, a loop that sends 25K-token prompts can make about 10 calls a minute before it hits 429s. Cached tokens do not count toward the limit, so a stable prompt prefix also buys headroom.
Groq Cost Examples
| Model | List price | With 60% of input cached | Batch (50% off) |
|---|---|---|---|
| gpt-oss-120b | $10.50 | $8.25 | $5.25 |
| gpt-oss-20b | $5.25 | $4.13 | $2.63 |
| Qwen3.8-27B (preview) | $60.00 | $48.00 | $30.00 |
The cached column assumes 30M of the 50M input tokens hit the cache at half price. A single request with 1M input and 200K output tokens on gpt-oss-120b costs $0.27, or $0.135 through Batch. An hour of audio costs $0.111 on Whisper Large V3 and $0.04 on Whisper Large V3 Turbo.
Groq vs Cerebras Pricing for gpt-oss-120b
Both providers serve gpt-oss-120b on custom silicon. Groq lists $0.15 input and $0.60 output per million tokens. Cerebras lists $0.35 and $0.75 (see Cerebras pricing). On the 50M input and 5M output workload above, Groq costs $10.50 a day at list price and Cerebras costs $21.25.
| Provider | Input | Output | Daily cost (50M in / 5M out) |
|---|---|---|---|
| Groq | $0.15 | $0.60 | $10.50 |
| Cerebras | $0.35 | $0.75 | $21.25 |
Price is one input. Rate limits, the models each catalog carries, and whether a provider offers batch or caching discounts matter as much for agent workloads. For more providers side by side, see LLM API pricing, Fireworks AI pricing, and Together AI pricing. Groq is not xAI's Grok; that naming clash is covered in Groq vs Grok.
Frequently Asked Questions
How much does the Groq API cost?
Per the GroqCloud models docs (October 2026), gpt-oss-120b costs $0.15 per million input tokens and $0.60 per million output tokens, gpt-oss-20b costs $0.075 and $0.30, and the preview model Qwen3.8-27B costs $0.80 and $4.00. Whisper Large V3 costs $0.111 per hour of audio and Whisper Large V3 Turbo costs $0.04 per hour.
Where did the Groq pricing page go?
groq.com/pricing now redirects to the Groq homepage. The per-model rates, speeds, context windows, and Developer plan rate limits are published in the models page of the GroqCloud documentation at console.groq.com/docs/models.
Does Groq have a batch discount?
Yes. The Groq Batch API costs 50% less than synchronous requests. You set a processing window between 24 hours and 7 days, and batch jobs do not count against your standard rate limits. Batch covers chat completions (text and vision) and audio transcription and translation.
Does Groq support prompt caching?
Yes. Prompt caching runs automatically on supported models with no code changes and no extra fee. When a request shares a prefix with a recent one, the cached portion is billed at 50% off. Cached tokens do not count toward rate limits, and cached data lives only in volatile memory and expires on its own.
What are the Groq rate limits?
The GroqCloud docs list Developer plan limits per model. gpt-oss-120b and gpt-oss-20b get 250K tokens per minute and 1K requests per minute. Whisper Large V3 gets 200K audio seconds per hour and 300 requests per minute. The Developer plan also unlocks the Batch API and Flex processing.
Is Groq cheaper than Cerebras for gpt-oss-120b?
On list price, yes. Groq charges $0.15 input and $0.60 output per million tokens for gpt-oss-120b. Cerebras charges $0.35 input and $0.75 output. A workload of 50M input and 5M output tokens a day costs $10.50 on Groq and $21.25 on Cerebras before any caching or batch discount.
The fastest endpoints are private deployments
Morph's top speeds come from dedicated deployments, not shared public endpoints: speculators trained on your traffic, caching tuned to your workload, and volume discounts over public per-token rates. Over 500 billion tokens per day run this way.
Build faster coding agents with Morph
Morph Router classifies each prompt and tells you which model to call; you keep your own provider keys. Pair Groq for cheap, fast gpt-oss steps with Morph Compact to cut context size by 50-70% before every call.