OpenRouter and Groq solve different problems at the cheap end of the market. OpenRouter is a router that resells hundreds of models at the provider’s list price and charges a 5.5% fee when you buy credits. Groq is a single provider running a short list of open-weight models on its own hardware, with GPT-OSS 120B at $0.15 per million input and $0.60 output and a 50% batch discount. Both offer something free, and the free tiers are where most readers start.
The trade-off is breadth versus depth. OpenRouter’s Free plan lists more than 25 free models capped at 50 requests a day until you’ve bought $10 of credits, after which the cap rises to 1,000. Groq’s Free plan exists but its numbers aren’t published; the only printed figures are Developer plan base limits such as 30 requests per minute on GPT-OSS 120B. Everything here was checked on both providers’ own pages on November 2, 2026.
Free models and the ten dollar trick
OpenRouter’s free models carry a :free suffix and the openrouter/free router picks one for you automatically. They’re limited to 20 requests per minute and 50 requests per day for accounts that have bought fewer than 10 credits, rising to 1,000 per day once lifetime purchases reach 10. Failed requests count toward the cap, and a negative balance blocks free models until you top up. Every new user also gets a small free allowance to test the platform. Spend $10 once and the daily ceiling is twenty times higher for the life of the account.
Groq offers a Free plan and a Developer plan, and downgrading returns you to Free tier rate limits and restrictions. The published rate limit tables, though, print only Developer base limits, and your organization’s exact numbers live on the limits page in account settings. Groq doesn’t publish Free plan figures anywhere official, so treat any third-party table as a snapshot that may already be stale.
Where the fee hides and what tokens cost
OpenRouter puts no markup on inference. You pay the provider’s list price for each model, and the fee lands on the purchase instead: 5.5% on the Standard plan and 8% on Business, charged when you buy credits. Bring your own provider keys and the first $25,000 of list-price inference each month carries no fee, then 5% applies. Opting in to prompt logging takes 1% off usage. On a $100 top-up, that 5.5% is $5.50 you never see again.
Groq prices its own catalog. GPT-OSS 120B is $0.15 input and $0.60 output with a 131,072 token context at roughly 500 tokens per second, GPT-OSS 20B is $0.075 and $0.30, and Qwen 3.8 27B is $0.80 and $4. The Llama 3.3 70B and 3.1 8B rows now read contact sales, so the Llama line is enterprise only. Whisper Large v3 Turbo transcribes at $0.04 per hour, which makes an hour of podcast audio cheaper than a coffee.
Groq stacks discounts OpenRouter can’t. Batch gives 50% off synchronous pricing with a window from 24 hours to 7 days, and prompt caching cuts cached input by 50% on the GPT-OSS models. Flex processing keeps on-demand pricing but allows 10 times the rate limits for paid customers, failing fast with a 498 when capacity runs out. Stack batch on an already cheap model and you’re paying $0.075 per million input on GPT-OSS 120B.
How hard you can push a paid plan
OpenRouter applies limits globally per account rather than per key, and spreading load across models helps because limits differ by model. Paid requests also face an in-flight spending budget, a fraction of your balance held against estimated cost, so a thin balance can trigger 402 errors even when it’s positive. Keep more in the account than you think you need. Budgets set by workspace admins reset at 00:00 UTC.
Groq’s Developer plan base limits are modest. GPT-OSS 120B and 20B get 30 requests per minute, 1,000 requests per day, 8,000 tokens per minute and 200,000 tokens per day, and Qwen 3.8 27B gets the same figures. Thirty requests a minute is one every two seconds, fine for a tool, tight for a chat product. Higher limits are available for select workloads and enterprise. Whisper models get 20 requests per minute and 28,800 audio seconds per day.
Prepaid balance or monthly invoice
OpenRouter is prepaid. Credits are US dollar deposits, with a $5 minimum and $25,000 maximum per transaction, and refunds for unused credits are only possible within 24 hours of purchase. OpenRouter reserves the right to expire unused credits 365 days after purchase, platform fees are non-refundable, and crypto payments are never refundable. Auto top-up is optional. Buy what you’ll use in a year, not more.
Groq’s Developer plan is postpaid. Upgrading costs nothing immediately, invoices arrive in arrears on a monthly cycle, and new customers hit progressive thresholds of $1, $10, $100, $500 and $1,000 before settling into monthly billing. A valid card, US bank account or SEPA debit is required, Groq bills only once usage reaches $0.50, and promotional credits appear mainly at hackathons and events.
Who sees your prompts
OpenRouter logs request metadata but not prompts or completions by default. Each upstream provider keeps its own retention policy, and your privacy settings can block routing to providers that may train on data, with separate toggles for paid and free models. Business plan accounts can pin requests to EU or US in-region routing. Free models typically come from providers whose policies you should read before sending anything sensitive. We’d keep customer data off the :free models entirely.
Groq doesn’t retain customer data for inference requests by default, logs inputs and outputs for up to 30 days only to troubleshoot reliability or investigate abuse, and lets any customer enable zero data retention in Data Controls. Batch files are kept up to 30 days. Prompt cache data lives in volatile memory and expires within hours.
Startup and event credits
Neither has an open startup application today. Groq’s service credits expire one year after purchase or issuance and can’t be transferred, sold or traded, and its promotions show up at hackathons rather than through a standing program. OpenRouter offers enterprise fee discounts by contract but no startup credit. Watch each provider’s own pages, since event credits appear and vanish within weeks.
The winner: OpenRouter
For most developers OpenRouter’s single prepaid balance, no inference markup and 25 plus free models beat Groq’s narrower catalog, even after the 5.5% purchase fee. Groq wins on pure cost when GPT-OSS fits, since batch and caching halve an already cheap rate, and on postpaid billing for teams that dislike prepaying. If your model list is short, Groq is the cheaper bill.
Choose OpenRouter if you want to switch between frontier and open models without new accounts, need free models for testing, or want routing that respects data policies.
OpenRouter offers (5)Choose Groq if GPT-OSS or Qwen serves your use case, you want batch and cache discounts on sub-dollar pricing, and you prefer a monthly invoice over prepaid credits.
Groq offers (4)