Groq API pricing: every Groq model, per 1M tokens

Separate Groq model usage from service-tier choices before estimating a workload. This page explains the charge categories and provides editable scenarios using current sourced records.

How Groq pricing works

Text models distinguish input from generated output. Supported cached input has its own treatment, and service tiers may use a different rate. Audio services use their documented media units. Live rows preserve these categories rather than blend them into one provider-wide quote. Official documentation.

Write down the unit before comparing alternatives. An input-heavy document workflow and a long coding answer can have very different cost patterns. For a tool-assisted system, identify hosted-tool charges before treating a token estimate as the full bill.

Full price list

The catalog and pricing reference are the starting points for supported identifiers and product rates. Review each source and verification date when a production budget depends on it. Official documentation.

An unavailable field means this reference does not have a verified amount for that combination. It does not mean the operation is free. Keep a list of excluded charges beside your estimate.

Estimate your cost

Use observed input and output from a representative request. The defaults are illustrative assumptions, so replace them before making a budget decision. Increase the traffic estimate to reflect the busiest plausible day and inspect its monthly implication.

Three worked scenarios

Chatbot. Include retained conversation and the system instruction in input. Compare a compact answer with the response your product actually needs.

Summarization. Start with the relevant extracted text and a defined output format. Repeated preprocessing or resending the source to several prompts can dominate the workload.

Coding assistant. Include repository context, error logs and follow-up attempts. Compare the cost of a completed task as well as a single call.

How to reduce cost on Groq

Prompt caching uses exact matching prefixes on supported models. Put stable instructions and definitions before changing request content, then inspect actual cache usage. Similar subject matter alone does not establish a cache hit. Official documentation.

Batch processing is an option for work that can be collected and completed asynchronously. Use it only when your application can reconcile individual results and tolerate the documented completion behavior. Official documentation.

Evaluate a less expensive candidate on saved tasks before switching. A lower request price can be offset by repair passes, unsupported answers or manual review. Record task outcomes with usage so a cost reduction remains meaningful after a future prompt change.

Billing, credits and limits notes

Developer billing can issue invoices during a cycle as progressive usage thresholds are crossed. This is separate from a spend limit. Monitor usage and payment state rather than infer remaining capacity from the most recent invoice. Official documentation.

Groq Spend Limits apply across the paid organization’s API keys. They are separate from project throughput controls. Test the application’s stopped-work behavior before raising the organization budget. Official documentation.

Read the Free plan guide if staying unpaid, or Groq throughput limits when a budgeted workload is throttled.

Price change history

Recorded changes describe observations retained by this site. For an invoice dispute or an older unrecorded interval, use provider billing records. Keep a date-stamped version of an estimate when approving a project.

Use the AI API cost calculator to turn the model and workload you are considering into an estimate.

Last verified 12 Sep 2026 · Source ↗

Frequently asked questions

Sources

• Price ↗
• Cache ↗
• Batch ↗
• Billing ↗
• Spend ↗

Last verified 12 Sep 2026 · Source ↗