Summary: Groq's Free plan gives you a real, no-card API key with daily caps that are generous for prototypes but tight for anything with users: on GPT-OSS 120B it's 30 requests a minute, 1,000 a day and 8K tokens a minute. The pay-as-you-go Developer plan lifts those caps sharply, keeps per-token prices low and adds Batch, Flex and spend limits. This breakdown shows exactly where the free limits bite, what the same work costs once you pay, and how Groq compares with Cerebras, the other fast-inference API people usually weigh it against.
What you actually get on the Free plan
You sign up at console.groq.com, open API Keys, create a key and start calling the API. A payment method only comes into play when you upgrade to the Developer tier, so the Free plan is genuinely free rather than a trial that converts.
The catch is the limits. Groq measures usage in requests per minute (RPM), requests per day (RPD), tokens per minute (TPM) and tokens per day (TPD), plus audio seconds per hour and per day (ASH/ASD) for speech-to-text. Whichever ceiling you hit first stops you. These are the base Free plan numbers on Groq's rate-limits page today for the models most people use:
- openai/gpt-oss-120b and openai/gpt-oss-20b: 30 RPM, 1K RPD, 8K TPM, 200K TPD.
- qwen/qwen3.8-27b: 30 RPM, 1K RPD, 8K TPM, 200K TPD.
- openai/gpt-oss-safeguard-20b: 3 RPM, 1K RPD, 2K TPM, 200K TPD.
- whisper-large-v3 and whisper-large-v3-turbo: 20 RPM, 2K RPD, 7.2K audio seconds per hour (two hours of audio) and 28.8K per day (eight hours).
- Orpheus text-to-speech (English and Arabic Saudi): 10 RPM, 100 RPD, 1.2K TPM, 3.6K TPD.
Limits apply to your whole organization, not to each person or key. Creating a second key for a teammate doesn't give you a second allowance.
Where the free caps bite first
For chat-style work the binding limit is usually TPM, not requests. Prompt and completion tokens both count, so 8K tokens a minute disappears quickly once you send long system prompts, retrieved documents or chat history. A single request carrying a 6K-token context leaves room for very little else in that minute, even though you've used only one of your 30 requests.
The daily 200K-token budget is the second wall. It's plenty for testing a feature, running a hobby bot or wiring up an agent framework, but a small app with real users will hit it on a busy day. Speech is more forgiving: eight hours of audio a day is enough to transcribe a decent backlog of meetings or podcast episodes for free.
One thing that stretches the free tier: prompt caching
Groq caches shared prompt prefixes automatically on the GPT-OSS models, and cached tokens don't count toward your rate limits. Keep the stable part of your prompt (system instructions, tool definitions, reference text) at the start and identical between calls, and put the changing user input at the end. Groq notes that cache hits aren't guaranteed and that cached tokens are still subtracted after processing, so heavy parallel requests can still hit the limit.
How to read your limits and stop guessing
- Check your organization's real numbers. The docs table is a summary with exceptions. Your exact values are on Settings → Limits in the console. If you hover over TPM and see an "X in / Y out" breakdown, your organization also has separate input and output token caps.
- Read the response headers. Every response includes
x-ratelimit-remaining-requestsandx-ratelimit-remaining-tokens. A detail that trips people up: the requests headers refer to your daily request limit, while the tokens headers refer to your per-minute token limit. - Handle 429 properly. When you exceed a limit, the API returns
429 Too Many Requests. Theretry-afterheader (in seconds) is only set on that 429, so wait at least that long before retrying instead of hammering the endpoint. - Watch usage. Dashboard → Usage shows consumption by model, which tells you whether one feature is eating your allowance.
The model-name gotcha in 2026
If you're following an older tutorial, check the model ID before blaming rate limits. Groq's deprecations page shows that llama-3.1-8b-instant and llama-3.3-70b-versatile were shut down on August 16, 2026 for free and Developer-tier usage. The suggested replacements are openai/gpt-oss-20b for Llama 3.1 8B and openai/gpt-oss-120b for Llama 3.3 70B. (The page originally pointed to qwen/qwen3.6-27b as the other option, but that model has since been deprecated in favor of qwen/qwen3.8-27b.) On the Models page both Llama models are now listed as Enterprise-only.
The confusing part is that the Quickstart code samples I checked on 2026-11-01 still use llama-3.3-70b-versatile. On a free or Developer account, swap that string for openai/gpt-oss-120b before running the sample.
What the Developer plan costs
Groq doesn't sell monthly or annual subscriptions for the API, so there's no billing-period toggle to worry about. The Developer plan is pure pay-as-you-go: you add a card, US bank account or SEPA debit account and pay per token or per hour of audio. Upgrading costs nothing up front. New accounts are invoiced when cumulative usage crosses $1, $10, $100, $500 and $1,000; after that you get one monthly invoice. Groq doesn't charge for amounts under $0.50, and you can downgrade back to Free from Settings → Billing after paying any outstanding usage.
Current list prices per million tokens from Groq's Models page:
- GPT-OSS 120B: $0.15 input, $0.60 output.
- GPT-OSS 20B: $0.075 input, $0.30 output.
- Qwen3.8 27B: $0.80 input, $4.00 output.
- Whisper Large V3: $0.111 per hour of audio; Whisper Large V3 Turbo: $0.04 per hour.
To put that in perspective, a day that uses 1 million input tokens and 200K output tokens on GPT-OSS 120B comes to $0.15 + $0.12 = $0.27 at list price. That's five times the Free plan's daily token budget on that model. Cached input tokens get a 50% discount, and so does the Batch API (the two discounts don't stack).
The Developer plan's base limits for GPT-OSS 120B jump to 1K RPM, 500K RPD and 250K TPM with no daily token cap listed. Whisper Large V3 goes to 300 RPM and 200K audio seconds per hour.
Three Developer-only tools worth knowing
- Flex tier: add
"service_tier": "flex"to a request for 10x the on-demand rate limits at the same price. The trade-off is that, when flex capacity runs out, requests fail fast with status498andcapacity_exceeded, so you need jittered backoff and retries."auto"lets Groq pick the best tier available to you. - Batch API: 50% cheaper, with a processing window you set from 24 hours to 7 days, and it doesn't touch your normal rate limits. Good for evaluations, bulk classification and backfills.
- Spend limits: under Settings → Billing → Limits, click Add Limit, enter a monthly USD budget and add email alerts (Groq suggests 50%, 75% and 90%). Only an organization owner on a paid tier can set this.
Groq vs Cerebras: free allowance and paid tiers
Cerebras is the obvious alternative because it also sells very fast inference for GPT-OSS 120B. The big difference is the free side: Cerebras says plainly that it has no permanently free tier. Prices below are list prices from each company's official pages on 2026-11-01. Both are usage-based, with no monthly or annual plan.
| Option | Cost to start | GPT-OSS 120B limits | GPT-OSS 120B price (per 1M tokens) | Worth knowing |
|---|---|---|---|---|
| Groq Free | $0, no payment method needed | 30 RPM, 1K RPD, 8K TPM, 200K TPD | Free within limits | Doesn't expire; no Batch, Flex or spend limits |
| Groq Developer | Pay as you go, billed in arrears | 1K RPM, 500K RPD, 250K TPM | $0.15 in / $0.60 out | Flex (10x limits), Batch (-50%), spend limits, chat support |
| Cerebras Free Trial | One-time $5 credit after adding a payment method | 5 RPM, 30K uncached TPM, 90K total TPM, 1M TPH, 1M TPD | Paid from the $5 credit | Credit expires after 30 days; access pauses when it runs out |
| Cerebras Developer | Buy pay-as-you-go credits | 1K RPM, 1M uncached TPM, 3M total TPM, no hourly/daily caps | $0.35 in / $0.75 out | Labeled for development and evaluation, not production |
In short: Groq is cheaper per token on GPT-OSS 120B and the only one of the two that's free indefinitely. Cerebras gives far more tokens per minute on its paid tier and counts only uncached tokens against its main limit. Also note that the model IDs differ: Groq uses qwen/qwen3.8-27b, while Cerebras uses qwen-3.8-27b.
Editor's note
Good fit: developers prototyping chat features, agents or transcription who want a free key that doesn't expire, and small projects that can live within 1,000 requests and 200K tokens a day.
Not a good fit: anything user-facing with unpredictable traffic on the Free plan (one busy hour can lock you out until the minute or day resets), or teams that need Llama models without an Enterprise contract.
Common errors and fixes:
429 Too Many Requests: you've hit RPM, RPD, TPM or TPD. Readretry-after, trim long prompts (TPM is usually the culprit) and keep stable prompt prefixes so caching helps.- Requests fail with an old model name: Llama 3.1 8B and 3.3 70B are gone for free and Developer accounts. Use
openai/gpt-oss-20boropenai/gpt-oss-120b. 498 capacity_exceeded: only happens on the Flex tier. Retry with backoff or drop back toon_demand.400withblocked_api_access: you've reached your own spend limit. Raise or remove it on the Limits page, or wait for the reset on the 1st of the month. Spend tracking lags 10 to 15 minutes, so you can slightly overshoot the cap.- Can't find the spend-limits page: it doesn't exist on the Free plan and needs organization-owner permission.
Server errors (500, 502, 503) aren't charged, so retrying them doesn't cost extra. You'll find screenshots of the official Groq and Cerebras pages I used at the end of this post.
FAQ
Is the Groq API really free?
Yes. The Free plan needs no payment method and doesn't expire, but every model has per-minute and per-day caps. You pay only if you upgrade to the Developer plan, and then only for what you use.
Can I use the OpenAI SDK with Groq?
Yes. Groq's API is mostly OpenAI-compatible. Point the client's base_url at https://api.groq.com/openai/v1, pass your Groq key, and use a Groq model ID such as openai/gpt-oss-120b. A few OpenAI features aren't supported, so check the compatibility page if a parameter is rejected.
When should I upgrade from Free to Developer?
Upgrade when you regularly hit the daily token cap, need more than 8K tokens a minute for long contexts, or want Batch processing and a spend limit. With no upfront fee and a low per-token price, the main reason to wait is if you don't want a card on file yet.
Related on lightmind24: OpenRouter Free Models: Limits, Fees and When to Pay, Tavily vs Exa vs Brave Search API: Free Tiers and Real Costs, and Deepgram vs AssemblyAI vs Gladia: Speech-to-Text API Costs.
Start on the Free plan, watch your token headers, and upgrade the day the daily cap starts shaping your product decisions.
Images: screenshots of Groq's public documentation (console.groq.com/docs) and Cerebras' public pricing page (cerebras.ai), captured by lightmind24.
Last verified: 2026-11-01
Sources: Groq Rate Limits, Groq Models, Groq Billing FAQs, Groq Spend Limits, Groq Flex Processing, Groq Service Tiers, Groq Batch API, Groq Prompt Caching, Groq Deprecations, Groq Quickstart, Groq OpenAI Compatibility, Groq Error Codes, Cerebras Rate Limits, Cerebras Pricing.
댓글
댓글 쓰기