Groq API Pricing in 2026: Every Model Priced, Verified

Groq's public rate card is now the shortest among major inference hosts: two paid text models, two paid audio models, and everything else behind an enterprise quote. The two that remain priced — OpenAI's gpt-oss-120b at $0.15 input / $0.60 output per million tokens and gpt-oss-20b at $0.075 / $0.30 — both cost roughly 2.3 times what Groq charged for the same checkpoints in June 2026, and every rate below is dated 2026-10-27 from groq.com/pricing.

Key facts

• Groq prices gpt-oss-120b at $0.15 input / $0.60 output per million tokens and gpt-oss-20b at $0.075 / $0.30 (as of 2026-10-27).
• Both paid text models are up roughly 2.3x from June 2026, when the same models were listed at $0.07/$0.30 and $0.05/$0.08 respectively (artificialanalysis.ai, ampm-aiops.com).
• Llama 3.1 8B, Llama 3.3 70B, Muse Spark Open, and MiniMax M2.7 are now enterprise-only with no published token price (groq.com/pricing).
• Prompt caching is live on all three GPT-OSS models: 50% off cached input tokens ($0.075/M on gpt-oss-120b), 2-hour expiry (docs).
• On gpt-oss-120b (high), Artificial Analysis's 72-hour medians measure Groq at 473.2 output tokens/second — third behind Cerebras (1,693.7) and SambaNova (701.2) — citing data as of late October 2026 (artificialanalysis.ai).
• Groq raised a $350M Series A at a $3.5B valuation on August 17, 2026, on top of $650M raised in June, as it pivots from LPU chips to operating an NVIDIA-based neocloud (groq.com, TechCrunch).

What happened

Groq's catalog has thinned and repriced at the same time. The current rate card splits every hosted model into three bands. Production models: gpt-oss-120b ($0.15/$0.60, listed at ~500 tokens/second), gpt-oss-20b ($0.075/$0.30, ~1,000 tokens/second), Whisper large-v3 ($0.111/audio hour) and Whisper large-v3-turbo ($0.04/audio hour). Preview models: Canopy Labs Orpheus TTS ($22–$40 per million characters), Llama Prompt Guard 2 22M/86M ($0.03/$0.04 per million tokens), the safety-tuned gpt-oss-safeguard-20b ($0.075/$0.30), and Qwen3.8-27B at $0.80 input / $4.00 output. Enterprise: Llama 3.1 8B, Llama 3.3 70B, Meta's Muse Spark Open, and MiniMax M2.7 — each shows "Contact Sales" where a price used to be.

That last band is the change with the most downstream effects. Llama 3.1 8B and Llama 3.3 70B — the two model IDs in nearly every Groq tutorial — were pulled from free and Developer accounts on August 16, 2026, the same day the deprecation page logged them moving to enterprise-only. Kimi K2 instruct, Qwen3-32B, and Llama 4 Scout had already been deprecated earlier in 2026. What is left on the self-serve card is OpenAI's open-weight GPT-OSS family plus a long tail of audio, guard, and TTS models.

The price moves compound the catalog shrink. Blended-price third-party data puts gpt-oss-120b at a 7:2:1 blended $0.14 per million tokens on Groq — mid-pack across the 16 providers benchmarked, where CoreWeave ($0.04), DeepInfra ($0.05), and Novita ($0.07) blend cheaper. Archive.org snapshots and third-party pricing guides show the same two models listed at $0.07/$0.30 and $0.05/$0.08 in June 2026. Groq has not published a changelog entry explaining the increase; we have not seen one, and the card itself carries no annotation.

Why it matters

Groq built its reputation on being the fastest place to run open models, and speed is still where it ranks well on its own turf — gpt-oss-20b's listed ~1,000 tokens/second and gpt-oss-120b's ~500 tokens/second on the card remain high for single-stream serving. But on the one model third parties benchmark across providers, Groq's 72-hour median of 473.2 output tokens/second now sits behind Cerebras (1,693.7) and SambaNova (701.2), with a 4.84-second time-to-first-answer-token that trails both. Its Endpoint Accuracy Index composite (86.36) carries a footnote: roughly 7% of tool-use tasks fail because Groq's API rejects function calls whose arguments do not exactly match the tool's schema.

The pricing move reads differently against the company's own announced pivot. After NVIDIA's December 2025 licensing deal absorbed founder Jonathan Ross and other LPU leadership, Groq LLC reconstituted as an NVIDIA Cloud Partner operating GPU datacenters — 13 of them, scaling from 54MW toward 200+MW in 2027, funded by a $650M June round and the $350M Series A in August. A rate card that steers commodity open-model traffic toward enterprise quotes, while the self-serve shelf narrows to two GPT-OSS checkpoints, is consistent with a business repositioning upmarket. Builders who priced Groq into their unit economics in June should re-run the math: the models most tutorials default to are no longer self-serve at any price.

Background

Groq entered 2026 as the LPU specialist: custom silicon, single-model latency leadership, a generous free tier, and a catalog that spanned Llama, Qwen, Kimi, and DeepSeek distills. The deprecation log now reads as a retreat from breadth — Kimi K2 instruct (March), Llama 4 Maverick (February), Llama 4 Scout and Qwen3-32B (June–July), the two Llama workhorses (August 16) — with replacement pointers that mostly lead back to GPT-OSS. Deprecations follow a stated process (email notice, transition window, then automatic upgrade), so existing integrations get warning, but the direction is one-way.

On privacy, Groq's Your Data page states that by default inference requests are not retained; exceptions are opt-in features that require persistence (batch files and fine-tuning data, retained 30 days or until deleted) and abuse-monitoring logs of up to 30 days, with a Zero Data Retention toggle all customers can enable. Retained data sits in US GCP buckets. These are Groq's own policy statements, not independent attestations — we flag them as claims, per our trust methodology.

For how the rest of the fast-inference field prices the same open models, see Cerebras Inference Pricing in 2026, SambaNova Cloud in 2026, and DeepInfra Pricing in 2026; for the wider field including routers and marketplaces, see our Groq alternatives roundup.

The full rate card (2026-10-27)

Rate limits are the Developer plan's base limits; Groq notes exceptions exist and exact per-organization limits live in the console's limits page. Cached tokens don't count toward TPM. Free-tier caps are published only inside the console; third-party guides through 2026 describe roughly 30 RPM / 6,000 TPM with no credit card required (cloudzero.com).

The cache math

Prompt caching on the GPT-OSS trio is automatic, free to enable, and cuts cached input tokens by 50% — $0.075/M on gpt-oss-120b, $0.0375/M on gpt-oss-20b — with prefixes expiring after 2 hours unused. At Groq's post-hike prices, the discount meaningfully softens the increase for agent-style workloads with long static system prompts; a 90% cache-hit workload on gpt-oss-120b blends to roughly $0.08/M combined. Artificial Analysis's cached rate for Groq confirms $0.075/M on cache hits at a measured 20% hit rate in its runs. Caching is not yet offered on Llama, Qwen, MiniMax, or Whisper — and with those text models now enterprise-only, the self-serve catalog and the cache-capable catalog are currently the same two models.

What 1M tokens costs on each paid text model

Compute for yourself with the card; the blended rows use standard mix ratios so cross-provider comparison stays apples-to-apples. Against the cheapest tracked hosts of the same model, Groq blends higher than CoreWeave, DeepInfra, and Novita on third-party medians — the premium buys its listed speed ceilings and OpenAI-compatible endpoint, not the lowest bill.

What's next

Three things will move this page. First, whether Qwen3.8-27B graduates from preview to a production price — its $0.80/$4.00 preview pricing is the only non-GPT-OSS text option self-serve customers have. Second, whether the enterprise-only band gets published pricing or stays quote-only; the more models that move behind Contact Sales, the less comparable Groq becomes in standardized price tables. Third, the neocloud build-out: Groq says it scales from 54MW to 200+MW in 2027 with NVIDIA hardware, and capacity additions could either restore a broad self-serve catalog or push pricing further upmarket. We re-check this card weekly as part of the Inference Price Index and will date every change.