Groq Pricing: Per-Model Rates, Free Tier, and Batch Discounts

Groq pricing runs per 1 million tokens and varies by model, from $0.05 per 1M input tokens on Llama 3.1 8B Instant up to $1.00 input and $3.00 output on Kimi K2. There is no flat subscription fee for API access. You pay only for the tokens you send and receive, and the rate depends entirely on which model you call.

Groq is not a model maker. It runs other labs' open models — Llama, Kimi, DeepSeek, and OpenAI's open-weight GPT-OSS models — on its own custom LPU chips, which is why the pricing sheet reads like a menu instead of a single number.

This guide breaks down the per-model rates, the free tier limits, the batch API discount, and what a real workload costs on Groq versus running the same job through GPT-5.6 or Claude.

How Groq Pricing Works

Groq charges per token, split into an input rate and a higher output rate, billed per 1 million tokens processed. Every model on GroqCloud has its own two-number price: what it costs to send text in, and what it costs to get text back out.

Output tokens always cost more than input tokens because generating text takes more compute than reading it. On most Groq models the output rate runs 1.3x to 4x the input rate, depending on the model.

There is no seat license, no per-user fee, and no minimum spend to start. A free account can call the API immediately, capped by the rate limits covered below.

• Pay-as-you-go, billed monthly by token volume
• No subscription tier gates model access — pricing gates only rate limits
• Prices are set by Groq per model, not by the model's original lab

Groq Price by Model (Per 1M Tokens)

Groq's cheapest model is Llama 3.1 8B Instant at $0.05 input and $0.08 output per 1M tokens, and its most expensive listed model is Kimi K2 at $1.00 input and $3.00 output. Between those two poles sit a mid-tier of Llama 4 and open-weight reasoning models.

• Llama 3.1 8B Instant: $0.05 input / $0.08 output — smallest, fastest, cheapest
• Llama 4 Scout: $0.11 input / $0.34 output — long-context, low-cost multimodal
• openai/gpt-oss-20b: $0.075 input / $0.30 output — OpenAI's smaller open-weight model
• openai/gpt-oss-120b: $0.15 input / $0.60 output — OpenAI's larger open-weight model
• Llama 4 Maverick: $0.50 input / $0.77 output — Meta's larger 2026 flagship open model
• Llama 3.3 70B Versatile: roughly $0.59 input / $0.79 output — the general-purpose default for most teams
• DeepSeek R1 Distill Llama 70B: roughly $0.75 input / $0.99 output — reasoning-style output at Groq speed, availability has shifted with DeepSeek model deprecations
• Kimi K2: $1.00 input / $3.00 output — Groq's priciest listed model, cached input drops to about $0.50

What the Groq Free Tier Includes

The Groq free tier gives every account API access with no credit card, capped at roughly 30 requests per minute and 500,000 tokens per day on a model like Llama 3.1 8B Instant. It is built for testing and light prototypes, not production traffic.

Limits differ by model — larger models on the free tier usually carry lower per-minute caps than small ones. Cached tokens do not count against the rate limit, which helps if your app repeats the same system prompt often.

• About 30 requests per minute, 14,400 requests per day on entry-level text models
• About 6,000 tokens per minute, 500,000 tokens per day on entry-level text models
• Whisper speech-to-text has separate free limits (around 20 requests per minute, 2,000 per day)
• No batch API or Flex processing on the free tier — those unlock on the Developer plan

Batch vs. On-Demand vs. Flex Pricing

Groq's Batch API cuts the standard token rate in half in exchange for asynchronous processing, while Flex processing costs the same as on-demand but grants roughly 10x higher rate limits during traffic spikes.

On-demand is the default: you send a request, you get a response in real time, you pay the full listed rate. Batch trades speed for cost — you upload a JSONL file of up to 50,000 requests, Groq processes it within 24 hours to 7 days, and every token bills at 50% off. The batch discount does not stack with prompt-caching discounts; batch tokens bill at the flat 50% rate regardless of cache status.

Flex processing is the opposite tradeoff. It keeps the standard price but removes the usual throughput ceiling, aimed at bursty production traffic rather than cost-sensitive background jobs. It is restricted to paid accounts.

• On-demand: full price, real-time response, standard rate limits
• Batch: 50% off, 24-hour to 7-day turnaround, best for nightly jobs and bulk classification
• Flex: standard price, ~10x higher rate limits, best for traffic spikes

Worked Cost Example: Groq vs. GPT-5.6 vs. Claude

A 10-million-input, 2-million-output-token workload costs roughly $7.48 on Groq's Llama 3.3 70B, versus about $44 on GPT-5.6 Terra and about $60 on Claude Sonnet 5 at standard rates — a 6x to 8x gap driven almost entirely by output-token pricing.

Here is the same math laid out model by model, using each vendor's per-1M-token rate at that volume.

• Groq, Llama 3.3 70B Versatile ($0.59 / $0.79): 10M x $0.59 + 2M x $0.79 = about $7.48
• Groq, Kimi K2 ($1.00 / $3.00): 10M x $1.00 + 2M x $3.00 = about $16.00
• OpenAI GPT-5.6 Terra ($2.00 / $12.00): 10M x $2.00 + 2M x $12.00 = about $44.00
• Claude Sonnet 5, standard rate ($3.00 / $15.00): 10M x $3.00 + 2M x $15.00 = about $60.00

Hidden Costs and Pricing Gotchas

The biggest hidden cost on Groq is model churn: preview and deprecated models can disappear or reprice with little notice, which breaks a hardcoded model ID in production. DeepSeek availability on Groq in particular has shifted more than once as Groq rotated which DeepSeek checkpoints it hosts.

The second gotcha is rate-limit-driven overage, not token overage. Free tier accounts hit the per-minute request cap long before they hit a dollar limit, and the fix is a paid plan upgrade, not a smaller prompt.

Third, blended "per million token" numbers you see quoted around the web assume a fixed input-to-output ratio. Your actual bill depends on your real ratio, and output-heavy workloads (long-form generation, code) cost more than the blended number suggests.

• Pin model IDs and monitor Groq's deprecation notices — do not assume a model ID is permanent
• Free tier fails on rate limits before it fails on cost — budget for a paid tier if you need production traffic
• Recompute cost from YOUR input/output ratio, not a vendor's blended average

When Groq Pricing Makes Sense

Groq pricing makes the most sense for high-volume, latency-sensitive workloads running open-weight models — chat interfaces, voice agents, and real-time classification where speed and per-token cost both matter more than frontier reasoning depth.

It makes less sense as your only vendor if your workload needs consistent multi-step reasoning, long-horizon agent planning, or the strongest available coding performance — those still lean toward GPT-5.6, Claude, or a comparably strong closed model.

• Good fit: real-time chat, voice agents, bulk classification, summarization, RAG retrieval formatting
• Weak fit: multi-step agentic coding, deep research synthesis, tasks where a wrong answer is expensive
• Common pattern: route easy, high-volume calls to Groq and escalate hard calls to a frontier model via a router like OpenRouter

Groq Alternatives and Where It Fits

Groq's closest competitors on speed-and-price are Together AI and Fireworks AI, both of which also host open-weight models on optimized inference hardware at similar per-token rates. The real choice usually comes down to model catalog breadth and regional latency, not price alone.

If your stack already routes across multiple providers, Groq slots in as one endpoint behind a router rather than a full replacement for a frontier-model subscription.

• Together AI — broader open-model catalog, similar pricing philosophy
• Fireworks AI — fine-tuning and custom deployment focus
• OpenRouter — routes across Groq and dozens of other providers from one API key