Guides / groq api cost
Groq API cost and the flattest output ratios anywhere
Groq API cost runs from $0.05 per million input tokens to $1.00, and output ratios go as flat as 1.34 times input on Llama 3.3 70B. That flatness breaks the usual advice to trim the answer first. Culpa, a local-first LLM cost, margin, and forecast ledger, prices each call from the model's own rates rather than a rule of thumb.
Why this happens
Two models in this book, Qwen3 32B and Llama 4 Scout, no longer appear on Groq's public pricing page as of 2026-08-01, so treat their rates here as historical until re-confirmed. Groq hosts open-weight models, and their price shapes differ from the hosted frontier line in a way that changes the optimisation. Llama 3.3 70B bills output at 1.34 times input, and Llama 3.1 8B at 1.6 times. On those models cutting your answer in half barely moves the total, because input and output cost nearly the same. Volume and prompt size become the levers instead, which is the reverse of every other provider here.
What this usually looks like
- You shortened responses and the bill barely responded.
- Input tokens dominate your Groq spend, unlike your other providers.
- A repeated system prompt costs full price, because five of the eight models record no cached rate.
- Cost per call is remarkably stable across requests, because response length barely matters.
Free, no card, no account
Run the free Cost Leak Scan
It shows your most expensive conversation before you install anything.
Mistakes that cost the most
| Mistake | Why it hurts | Do instead |
|---|---|---|
| Carrying output-trimming habits over from a frontier provider. | On a 1.34x ratio, halving the answer cuts a small fraction of the bill. | Work on prompt size and call volume instead. On these models that's where the money sits. |
| Assuming a cached rate exists. | Five of the eight Groq models here record no cached rate, so repeated prompts pay in full. | Check the model before budgeting a cache discount, and shorten the repeated block instead. |
| Treating all open-weight models as one price tier. | The spread is 20x on input, from $0.05 to $1.00, and ratios range from 1.34x to 5x. | Price each model against your own token split. The cheapest input isn't always the cheapest total. |
Run this check tonight
- Compute your own output-to-input token ratio and compare it against the model's price ratio.
- Check whether your Groq model records a cached rate at all before budgeting one.
- Measure the repeated share of your prompt. Without a cache, that block pays full price every call.
- Price the same workload across two or three Groq models. The ranking often flips with your token split.
- Confirm retry caps on high-volume Groq paths. Cheap models attract loose retry policies.
Why halving the answer saves almost nothing here
Illustrative example
10 million input and 4 million output tokens a month on Llama 3.3 70B at real rates, $0.59 per million input and $0.79 output.
Cutting the prompt saves nearly twice what cutting the answer saves on this model. That's the opposite of the advice that works on every frontier provider.
Every number, with its confidence and source
| Figure | What it means | Confidence | Source |
|---|---|---|---|
| 1.34x | flattest output-to-input ratio in this price book, on Llama 3.3 70B via Groq | calculated | $0.79 output divided by $0.59 input per million, from the price book, effective 2026-07-02. |
| 20x | spread between the cheapest and dearest Groq input rates | calculated | $1.00 per million on kimi-k2 divided by $0.05 on llama-3.1-8b-instant, from the price book. |
What a generic answer can’t know
A rate card can tell you the ratio is flat. It can't tell you your own input-to-output split, which is what decides whether that flatness helps or hurts you, or which feature sends the long prompts. Culpa records both on your infrastructure, keeps your prompts and responses there, and counts the calls to run your plan.
Questions founders ask next
Which Groq model has the flattest price ratio?
Llama 3.3 70B Versatile, at $0.59 input and $0.79 output per million, a ratio of 1.34. Llama 3.1 8B Instant is next at 1.6. On both, prompt size matters more than answer length.
Do Groq models support cached input?
Only three of the eight in this price book do. Llama 3.1 8B, Llama 3.3 70B, Llama 4 Scout, Qwen3 32B and Qwen3.6 27B all record no cached rate, so a repeated system prompt on those pays the full input rate every call.
Is the cheapest Groq model the cheapest overall?
Not necessarily. Llama 3.1 8B has the lowest input rate at $0.05 per million, but ratios vary from 1.34x to 5x across the line, so the cheapest total depends on your own input-to-output split.
On your infrastructure
Culpa runs on your infrastructure. Your prompts and responses never leave it. Culpa counts calls to run your plan, and it fails open, so if it ever breaks your app keeps running.
Free, no card, no account
Run the free Cost Leak Scan
It shows your most expensive conversation before you install anything.
Keep reading
Sources: Groq pricing. Last reviewed 2026-08-01, rates effective 2026-07-02. Plain text version.