Groq vs OpenRouter: which purchase route fits your workload?

Choose Groq direct when you want Groq’s serving and can manage that integration. Choose OpenRouter when access to other providers, routing controls or consolidated billing is useful. For the same GPT-OSS 120B route, the published base token rates matched in our check; an alternative OpenRouter provider had a different price and precision label.

The important distinction is between buying Groq directly, buying Groq through OpenRouter and buying another host’s deployment through OpenRouter. A lower model headline price does not establish a cheaper way to buy the same host.

Checked 8 November 2026 using public documentation and an endpoint response. Calculations below are budgets, with no paid inference or speed, quality and reliability benchmark.

Compare your token mix first

Use the same planning period for all three rows. Keep the billed output total, including reasoning tokens where applicable. Actual cache reads can differ by route, so each route has its own input in the expandable settings. An identical prompt is not proof that either service returned a cache hit.

What the three prices actually compare

All rates are USD per million tokens. The direct cached-input rate applies Groq’s documented 50% hit discount to its published input rate. The OpenRouter cache rates are taken from the named endpoint’s response.

Sources: Groq’s supported models and rates, Groq’s cache rules, and OpenRouter’s GPT-OSS 120B endpoint records. Rates are a dated observation, not a reservation or guarantee of availability.

The two Groq rows match on these base rates. PrimeIntellect is a separate deployment, labeled FP4 in the endpoint record; the cited Groq rate table does not specify its precision. That prevents this table from establishing equivalent output quality or application performance. Test the features and results you need before treating a different deployment as a substitute.

For an uncached period with 10 million input and 2 million output tokens, Groq’s model usage is $2.70. OpenRouter’s Groq route has the same usage subtotal; allocating a 5.5% Standard funding fee gives $2.8485. The illustrated PrimeIntellect route has $0.414 of model usage, or $0.4368 with the same proportional allowance. These numbers are generated from the same rates used in the calculator.

The alternative’s lower subtotal combines a different upstream offer with OpenRouter’s funding model. It is not a measurement of savings from changing only the gateway. Extra work to verify outputs, handle unsupported features or recover failures belongs in your eventual task budget.

Separate usage from paying for credits

OpenRouter’s current pricing page says Standard credits carry a 5.5% purchase fee, while Business uses 8%. The fee is charged when credits are bought. It is not a second token charge on every request.

The calculator allocates a percentage of usage for comparison. It does not simulate actual top-ups. For a cash budget, write down the credits you already own, the amount you will purchase, the checkout fee and any unused balance. A minimum purchase or minimum payment fee can make a small top-up differ from a simple percentage. Taxes, exchange costs and paid tools also need separate entries where they apply.

If your credits are already funded, the immediate cash payment for a request can be zero while its usage still consumes economic value. Keep those two views separate. Likewise, a free allowance or a negotiated contract changes the applicable budget; replace the fee allowance rather than adding another platform fee to an already inclusive quote. BYOK has separate rules and is outside this example.

Pin the provider when the comparison requires Groq

On OpenRouter, selecting a model is not the same as requiring a particular provider. Its provider routing documentation describes provider restrictions and fallback controls. An order expresses preference; a strict provider restriction and your fallback choice determine whether another provider can serve the request.

When checking the Groq-to-Groq rows, record the upstream that actually served each request. If fallback sends a request elsewhere, classify that bill under the other route. A more resilient route may be useful, but it answers a different purchasing question from a strict comparison of one host.

For direct integration, Groq documents an OpenAI-compatible endpoint. Compatibility still needs verification for your model, parameters, tools and response handling. Changing credentials and a base URL is only the beginning of a migration check.

Check cache reads and accepted results

Start the budget with zero cached input until you have usage evidence. Then enter each route’s actual cached share. A lower cache-read price can be outweighed by fewer hits or a different output volume. Avoid carrying one platform’s observed hit rate into the other row as if it were measured there.

Use a small set of tasks you can assess: for example, structured extraction with required fields or code changes with a known validation command. Keep the prompts, reasoning settings and output limits recorded. Measure completion time and count accepted outputs, failed attempts and billed tokens. Published throughput or network latency alone cannot establish the cost of finishing those tasks.

If all routes satisfy the task, compare their token budgets alongside setup and maintenance effort. If one fails a required feature or produces unacceptable results, its low token subtotal does not settle the choice. Our API versus rented GPU and hardware guide extends that decision to other ways of running the workload.

Questions before switching

Is Groq cheaper than OpenRouter?

It depends on the host and model you select. The Groq-to-Groq example has matching base token prices and separate funding costs. Another OpenRouter host can have a different rate and service conditions. Calculate the route you would actually purchase.

Does OpenRouter always use Groq for GPT-OSS?

No. A model can list several providers. Apply the current routing restrictions when Groq is a requirement, and verify the served route and any fallback in your usage records.

Can I compare speed with this calculator?

The calculator prices tokens and a fee allowance. It does not measure latency, output quality, concurrent capacity or uptime. Measure those for your own task before using the budget to select a route.

For another direct-versus-router example, continue with Chutes vs OpenRouter. For a subscription purchasing option, see NanoGPT vs OpenRouter.