Groq vs OpenRouter

When to pick Groq

• Open-weight inference is the whole workload and volume is high. Groq charges per token with no platform fee, so $0.05 per million input tokens on Llama 3.1 8B is what you actually pay — OpenRouter adds 5.5% on every credit purchase, which on a $10 top-up hits the $0.80 minimum and works out to 8%.
• Latency is the product. Groq publishes per-model throughput on its own pricing page — 1,000 tokens/sec on GPT-OSS 20B, 840 on Llama 3.1 8B — where OpenRouter publishes no first-party speed figure at all, only measured rates per provider endpoint.
• You have batch work or strict data rules. Groq’s batch API takes 50% off with a 24-hour to 7-day window and does not touch your standard rate limits; OpenRouter has no batch API. Groq also offers a zero-data-retention toggle on every account, not just enterprise ones.

When to pick OpenRouter

• You need models Groq does not have. Groq’s priced roster is six open-weight text models plus Whisper and TTS; OpenRouter returns 338 models from 57 authors, including Claude Opus 4.8 at $5/$25 per million and Gemini 3.5 Flash at $1.50/$9 — no closed frontier model is available on Groq at any price.
• Uptime matters more than the last few percent of cost. OpenRouter serves a single model from as many as 20 provider endpoints and fails over automatically on error or rate limit, with sticky routing to keep prompt caches warm. Groq is one provider, so its outage is your outage.
• You are still deciding. Because OpenRouter passes provider rates through without markup, you can benchmark the same model across providers on price and measured speed before committing — and one of those providers is Groq itself.

Bottom line

The honest framing is direct provider versus aggregator, not peer versus peer: Groq is a listed provider inside OpenRouter, and OpenRouter routes GPT-OSS 120B to Groq at $0.15/$0.60 — exactly Groq’s own posted rate. That makes “Groq through OpenRouter” a real third option, buying failover and one integration at the cost of the 5.5% credit fee. Going direct to Groq is right when your traffic is open-weight, steady, and latency-sensitive enough that published throughput and a 50%-off batch tier pay for the narrower catalog. OpenRouter is right when you need frontier models alongside cheap ones, or when a single provider’s outage is unacceptable. Two cautions on the numbers. OpenRouter markets “400+ models from 70+ providers” while its own public endpoint returns 338 models and 94 providers, so treat the headline as vendor copy rather than fact. And on OpenRouter your data policy is inherited from whichever provider serves the request — its own table flags DeepSeek and NVIDIA as “may train” — so enforce zero retention explicitly rather than assuming it.