Groq pricing in 2026: every model, free tier, and hidden discounts explained

What is Groq (and why does pricing work differently here)?

Groq doesn't make models - they run other people's models (Llama, Qwen, Mistral, Whisper, OpenAI open-weights) on their own custom silicon: the Language Processing Unit, or LPU. Founded in 2016 by former Google TPU engineers, they raised $750M at a $6.9B valuation in September 2025 and now serve 2M+ developers. The McLaren F1 team uses Groq for real-time race analytics - not a use case where "usually fast" is acceptable.

The pricing model is simple: charge per token, no idle infrastructure fees, no elastic pricing spikes. Groq's official statement on it: "Other inference providers spike costs without warning. Some hide behind elastic pricing. Groq pricing is linear and predictable, with no hidden costs or idle infrastructure."

Why the LPU changes the cost equation

GPUs were built for training: large external DRAM/HBM memory hierarchies, dynamic scheduling, cache coherency protocols. These are good properties when parallelizing matrix operations across thousands of cores for training. They're a mismatch for inference, where sequential layer execution has low arithmetic intensity and memory fetches dominate latency.

The LPU architecture takes a different approach. On-chip SRAM serves as the primary weight storage - not a cache, the primary memory. Groq's purpose-built compiler pre-schedules every operation down to individual clock cycles before execution starts, eliminating dynamic scheduling overhead entirely. The RealScale chip-to-chip protocol lets hundreds of LPUs behave as a single core for tensor parallelism. Because every operation is statically scheduled, Groq can run pipeline parallelism on top of tensor parallelism: layer N+1 begins processing while layer N is still finishing - something GPU dynamic scheduling can't reliably do.

The practical result: GPT OSS 20B at 1,000 tokens per second. Llama 3.1 8B at 560–840 TPS. Llama 3.3 70B at 280–394 TPS. Typical GPU-based cloud APIs run 50–100 TPS on equivalent models. When the same hardware serves more requests per second, fixed costs spread across more tokens - which is how $0.05 per 1M input tokens becomes commercially viable.

Groq free tier: what you actually get

The free tier requires no credit card and is governed by rate limits, not a monthly token budget. Here's exactly what each model provides on the free plan:

(RPM = requests per minute, TPM = tokens per minute. Source: Groq rate limits docs)

Two things catch developers off-guard here. First, rate limits apply at the organization level, not per API key. Creating five keys doesn't give you 150 RPM - it's still 30 RPM shared across your whole account. Second, prompt caching tokens don't count toward rate limits, which is a meaningful benefit if you have long system prompts that repeat across calls.

The per-minute TPM limits are usually the real constraint, not the daily request caps. A 2,000-token prompt eats a third of the Llama 8B TPM budget in one call.

"I've been using the Groq API nonstop, constantly thinking to myself 'how have I still not hit some kind of free tier limit'"

@ctatedev, May 2024

The Whisper free tier is the standout value. Artificial Analysis confirmed Groq as one of the lowest-cost Whisper Large v3 providers. On the free plan you get 2,000 audio transcription requests per day - roughly 2 hours of audio per clock-hour when batching at the 10-second minimum per request. OpenAI charges $0.36/hour for Whisper access; Groq's paid tier charges $0.04–$0.111/hour, so the free tier is a generous place to start.

"Their free API for speech to text is amazing, so generous, highly recommend."

Trustpilot reviewer, search-derived

Groq paid API pricing: every model

All prices are in USD per 1M tokens (input / output) unless otherwise noted. Source: Groq pricing page.

Text/LLM models

A few model notes worth pulling out. GPT OSS 20B - OpenAI's open-weight model, not GPT-4 - runs at 1,000 tokens per second at $0.075 input / $0.30 output. That is simultaneously the fastest model on the platform and one of the cheapest per output token. Llama 4 Scout supports vision inputs (up to 20 MB files) but remains in Preview - don't put it in production yet. Kimi K2 is the only model where prompt caching is explicitly baked into the pricing row: $0.50 per 1M cached input tokens versus $1.00 uncached.

The Prompt Guard models ($0.03–$0.04 per 1M tokens) are safety classifiers designed to detect prompt injection and jailbreak attempts - useful if you're building customer-facing AI and need a lightweight filter layer ahead of your main model.

Developer plan rate limits

The jump from free to Developer plan is substantial:

(Source: console.groq.com/docs/models)

How Groq pricing compares to OpenAI and other providers

The figure most commonly cited in developer communities is "10–20x cheaper than OpenAI for comparable open-source models." That's roughly accurate, with the necessary qualifier that you're not comparing identical models.

"LLM inference on Groq costs about 10 times less compared to OpenAI's pricing for GPT-4o. Groq is 10–20x cheaper, but for a somewhat less capable model - Llama 3-70B vs GPT-4o."

Balazs Kocsis, Medium, August 2024

The most honest comparison isn't Groq versus OpenAI's proprietary models - it's Groq versus other open-source hosting providers like Together AI or Fireworks AI running the same models. There, according to the Awesome Agents 8-week production review, Groq runs 20–50% cheaper at equivalent model tiers with deterministic tail latency that p99 stays within 15% of median - a meaningful advantage over GPU workloads where tail latency spikes are common.

"Goodbye OpenAI API. Today, you can get the same underlying intelligence - Llama-3 or its open-source competitors - for rates plummeting toward the floor, often below $0.20 per million tokens. That is a 99% price reduction in eighteen months."

Aparna Pradhan, GoPenAI, December 2025

The practitioner mental model that's emerged in the developer community - summarized by Jolly Gupta on LinkedIn (114 likes, September 2025): use Groq for speed-critical and cost-sensitive open-source workloads, use OpenAI when you need GPT-4o's capabilities or multimodal depth. Most production stacks do both.

Groq also appeared in the Artificial Analysis survey as one of the top-5 inference providers by developer adoption - alongside OpenAI, Google, Anthropic, and Microsoft.

Audio pricing: Whisper and TTS

Speech-to-text

Groq runs both Whisper Large v3 variants on LPU hardware, delivering transcription at 217–228x real-time speed. An hour of audio processes in about 15 seconds.

For most workloads, Turbo at $0.04/hour is the clear choice - faster and 2.8x cheaper than the full model, with only marginal quality differences on most audio. Audio is billed at a 10-second minimum per request regardless of actual length, so batching short clips together is worth the implementation effort.

OpenAI charges $0.36/hour for Whisper; Groq at $0.04/hour is 9x cheaper on the Turbo model. Levels.io noted that Whisper + TTS on Groq was "very cheap" even in 2024; the pricing has remained stable since.

Text-to-speech (Preview)

Groq recently launched TTS via Canopy Labs' Orpheus models:

These are still Preview status. The LPU speed advantage is visible here too - Orpheus generates at 100 characters per second on Groq, which enables near-real-time voice applications.

Compound AI systems: when tools cost extra

GroqCloud's Compound systems - groq/compound and groq/compound-mini - are agentic wrappers that give a language model built-in web search and code execution. Pricing is model token costs plus tool usage:

The Compound system runs at ~450 TPS with 131k context. It's a practical starting point for agentic AI workloads where you want to delegate tool-use orchestration to the platform rather than build it yourself.

Two hidden discounts worth knowing

Batch API: 50% off for async workloads

The Batch API halves the cost of any model by running jobs asynchronously. You submit a JSONL file (up to 50,000 lines, 200 MB), processing completes within 24 hours to 7 days, and you pay 50% of the standard per-token rate. No impact on your standard rate limits.

This is the right call for: document classification pipelines, bulk content generation, nightly data enrichment, content moderation at scale - anything where latency tolerance earns you a significant discount. Tool usage in Compound systems is still charged at standard rates.

Prompt caching: 50% off repeated prefixes

Prompt caching is automatic - no code changes, no extra fee. When the same prefix (a long system prompt, a reference document) repeats across calls, Groq caches it for up to 2 hours. Cache hits cost 50% of the normal input price.

Models supporting prompt caching and their cached rates:

The double benefit: cached tokens cost half as much and don't count toward rate limits. For workloads with long system prompts - RAG pipelines, document Q&A, AI customer support agents with large knowledge contexts - this meaningfully extends your effective throughput without upgrading your rate limit tier.

Rate limits: what happens when you hit them

When any rate limit is exceeded, Groq returns HTTP 429 with a retry-after header showing how many seconds to wait. The error body is specific:

"Rate limit reached for model openai/gpt-oss-20b … service tier: on_demand … Limit 200,000 · Used 199,336 · Requested 1,524 · Please try again in 6m 11.52s."

Standard Time project management tool documentation, April 2026

Response headers also include x-ratelimit-limit-requests, x-ratelimit-remaining-tokens, and x-ratelimit-reset-requests - enough to implement precise exponential backoff without trial and error.

The key operational consideration: rate limits are per organization, and per model. If you're running multiple services or team members from the same Groq account, they share the same limit pool. Use separate organization accounts for production and development environments, or contact Groq about higher limits for specific workloads via console.groq.com/settings/limits.

Enterprise pricing

There's no public enterprise pricing table. To access the following, contact groq.com/enterprise-access:

• Higher rate limits for specific workloads
• GroqRack on-premises deployment
• LoRA fine-tuned models
• Enterprise-only models (Minimax M2.5, Qwen3-VL 32B with vision)
• Regional deployment and data residency options
• SOC 2, GDPR, and HIPAA compliance documentation

On uptime: the Awesome Agents production review measured 99.94% uptime over 8 weeks with p99 latency within 15% of median - better tail behavior than GPU-based competitors because LPU scheduling is deterministic. Enterprise SLA guarantees require a formal agreement.

The sustainability question

Most Groq pricing guides skip this. We won't.

In September 2024, Kyle Corbitt posted on X that he'd overheard a Groq employee claim their per-token costs are "1–2 orders of magnitude higher than what they charge." The post hit 271k views. Earlier in 2024, @swyx did the math and found the pricing only works at a batch size of ~512 - unheard of in normal inference - and drops to ~$1.84 per million tokens at a normal batch of 64.

The counterargument: Groq raised $750M from BlackRock, Samsung, Cisco, and Disruptive AI specifically because the volume and new-chip thesis is credible. Their customer case studies show GPTZero at 7x faster and 50% lower cost, ReBlink at 14x lower cost per game, Recall at 10x lower cost. PeerSpot mindshare data shows a slight decline YoY (13.7% to 9.8%) among enterprise AI infrastructure evaluators, which may reflect NVIDIA deal uncertainty - worth monitoring.

Our take: we don't know if current pricing is structurally sustainable or a deliberate land-and-expand strategy ahead of second-generation chips. What we know is that pricing has been stable through 2025–2026 and the $750M raised buys time. Use it where the price-performance makes sense; don't architect yourself into a single-provider dependency you can't swap.

Who should (and shouldn't) use Groq

Use Groq when:

• You're building real-time voice or chat interfaces where 280–1,000 TPS matters to the user experience
• Your model stack runs on Llama, Qwen, Whisper, or OpenAI's open-weight models
• You need cheap transcription at scale - Whisper Turbo at $0.04/hour is hard to beat
• You're prototyping - the free tier covers most development workloads without a credit card
• You have async batch workloads - the 50% Batch API discount changes the economics significantly

Look elsewhere when:

• You need GPT-4o, Claude, or Gemini - not available on GroqCloud
• You need robust multimodal support - Llama 4 Scout is Preview only
• You need on-premises deployment with standard support terms - GroqRack requires enterprise negotiations
• You need fine-tuned proprietary models - LoRA fine-tuning requires enterprise access

For a broader feature comparison, our Groq review covers the full product in depth. If you're still weighing providers, Groq alternatives compares Together AI, Fireworks, Cerebras, and others across the same price-performance dimensions.

Try eesel for AI-powered customer support

If you're evaluating Groq for customer support or helpdesk automation, eesel pairs well with it. eesel deploys autonomous AI agents directly inside your existing tools - Zendesk, Freshdesk, Slack, email - and routes support tickets to the right model based on complexity. Simple, high-volume queries go to a fast, cheap model tier (exactly what Groq's Llama 8B and GPT OSS 20B are built for); complex escalations go to a higher-capability model.

Teams handling 100,000+ tickets per month use eesel agents that actually resolve issues rather than just deflecting them - no new interface to learn, no prompt engineering required. You brief the agent the way you'd onboard a new employee, and it handles the rest.