inference/research
verifiedhigh · 0.85updated 2026-10-04

Groq API token pricing (serverless LPU, per 1M tokens, 2026-10-04 list price)

Groq cloud (US); no regional GPU SKU to price

Fields

id
supply/rent-groq-api-token-pricing
Kind
cloud
Region
Groq cloud (US); no regional GPU SKU to price
Channels
serverlessapilist-price
Price (USD)
0.075
Price basis
USD per 1M TOKENS, LIST PRICE, observed 2026-10-04. openai/gpt-oss-20b $0.075 input / $0.30 output at 1000 tok/s; openai/gpt-oss-120b $0.15 / $0.60 at 500 tok/s; qwen/qwen3.8-27b $0.80 / $4.00 at 450 tok/s; openai/safety-gpt-oss-20b $0.075 / $0.30. Speech is priced by time not tokens: whisper-large-v3 $0.111 per hour, whisper-large-v3-turbo $0.04 per hour. Content-moderation models are priced per 1M tokens AND per 1M CHARACTERS on the same page (Canopy Labs Orpheus V1 English $22.00 per 1M characters, Arabic Saudi $40.00). price_usd = gpt-oss-20b input PERISHABLE-FIGURE STAMP (added 2026-10-04, uniform across data/supply/ so it is greppable). PRICE CLASS: MIXED - this record contains sales-gated cells as well as published ones. Every price here is an ASKING price a seller or provider advertises: NOT a sold, completed or realised transaction price and NOT a negotiated rate. A cell reading 'ContactSales' is not a price at all and deliberately carries no number. OBSERVED: 2026-10-04 RE-VERIFY BY: 2027-01-04 - the repo's quarterly re-verification bar. After that date this figure is presumed drifted and must be re-checked against its source before it is quoted again. A price without its observation date is noise; this stamp exists so a reader can tell a fresh figure from a stale one without reading prose. UNIT: USD per 1M TOKENS, LIST PRICE, observed 2026-10-04 READING RULE: No further perishable flag is published for this figure.
Availability
in-stock
Lead time source
structural-na
Lead time basis
lead_time_weeks is null because NO NUMBER APPLIES, and the reason is specific rather than generic: this record prices a metered service term (tokens, egress, storage) rather than hardware capacity, so there is no order-to-delivery interval to report at all. Calling this 'no published lead time' would be wrong in the other direction - nothing was withheld.
Export control regime
no EAR citation: no controlled item is transferred by this transaction; not an EAR citation: provider service terms and end-use policy (the AI Diffusion Rule briefly tiered service access before its 2025-05-13 rescission)
Export control note
MECHANISM: a token-priced API, so no accelerator crosses a border and the 3A090 hardware licence regime does not attach to the transaction. Export control becomes an END-USER AND TERMS question instead: the provider applies geographic and end-use restrictions to service access, and a customer can still be refused under sanctions or end-use policy wherever the compute sits. A reader targeting a restricted destination should read the provider's terms, not the EAR.
notes
Two things make Groq's page structurally different from the others. First, its headline Llama models are NOT on the public list price at all: llama-3.1-8b-instant and llama-3.3-70b-versatile both read literally 'ContactSales' in the price column. So the cheapest commonly-quoted Groq number in circulation is stale - Llama 3.3 70B on Groq moved to sales-gated pricing, and only the gpt-oss / Qwen / guard models remain self-serve. Second, Groq publishes SPEED next to price on the same row (1000 and 500 tok/s for the gpt-oss models), which makes it the only provider here where throughput is a published list attribute rather than something you benchmark yourself. For character-priced models, tokens are the wrong unit - $22 per 1M CHARACTERS on a TTS-style model is not comparable to a $2 per 1M token input rate at all. No GPU hourly rate is published. EXPORT CONTROL (2026-10-04): NOT APPLICABLE IN THE HARDWARE-SHIPMENT SENSE - this record prices a token-based API, so no accelerator is transferred across a border and the ECCN 3A090 hardware licence regime does not attach to the transaction. Recorded null rather than false precisely because a token API is where export control instead becomes an END-USER/TERMS question: the provider applies geographic and end-user restrictions to service access (the AI Diffusion Rule briefly put a tiered country structure on exactly this kind of service before being rescinded on 2025-05-13), and a customer can still be refused under sanctions or end-use policy regardless of where the compute sits. Anyone reading this record for a restricted destination should look at the provider's terms, not at the EAR. Sources for the regime: spl-bis-2025-05-ai-diffusion-rescission, spl-bis-2026-05-d5-guidance. EXPORT CONTROL (2026-10-04): NOT APPLICABLE IN THE HARDWARE-SHIPMENT SENSE - this record prices a token-based API, so no accelerator crosses a border and the ECCN 3A090 hardware licence regime does not attach to the transaction. Recorded NULL rather than false precisely because a token API is where export control instead becomes an END-USER and TERMS question: the provider applies geographic and end-user restrictions to service access (the AI Diffusion Rule briefly put a tiered country structure on exactly this kind of service before being rescinded and unenforced from 2025-05-13), and a customer can still be refused under sanctions or end-use policy regardless of where the compute sits. For a restricted destination, read the provider's terms, not the EAR. Sources: spl-bis-2025-05-ai-diffusion-rescission, spl-bis-2026-05-d5-guidance. LEAD TIME: null - correct and not a gap. A token-priced serverless API is provisioned on request, so there is no lead time to publish; the binding constraint when a model is at capacity is provider-side queuing and any per-account concurrency limit, which this repo records in notes where a vendor publishes one (see the DeepSeek concurrency figures already here). Deliberately NOT recorded: any 'model capacity lead time' estimate, because vendors publish none and the SEO 'GPU lead time' numbers that circulate do not apply to a rented endpoint. Probe log in source spl-negative-no-oem-publishes-lead-time. ACCELERATOR LINK ADDED 2026-10-04: Groq serves on its own LPU ([[accelerators/groq-lpu]]), not on a rented GPU, which is why Groq is the one provider here where published throughput (1000 tok/s gpt-oss-20b, 500 tok/s gpt-oss-120b) is a list attribute rather than something you benchmark - there is no commodity GPU behind the rate to be throttled by. VOLATILE STRUCTURE, RECORDED WHILE IT IS STILL TRUE - re-verified against the live model list 2026-10-04. THIS GATING PATTERN MOVES, and it is recorded here at a known date because it will not move in one direction. AS OF 2026-10-04, THREE rows on Groq's production and preview model tables read literally 'ContactSales' in BOTH the price column and the rate-limits column, and each is additionally tagged 'Enterprise' in the model list: (1) llama-3.1-8b-instant, 560 tok/s, 131,072 context; (2) llama-3.3-70b-versatile, 280 tok/s, 131,072 context, max completion 32,768; (3) minimaxai/minimax-m2.7, 260 tok/s, 196,608 context, max completion 131,072. Models 1 and 2 were already recorded; minimaxai/minimax-m2.7 was NOT in the previous revision of this record and is added by the 2026-10-04 re-verification. THE DIRECTION OF TRAVEL IS THE POINT, and it cuts against the intuitive reading. The FIRST Llama rows to move to ContactSales were Meta's own first-party open-weight models - the two most widely quoted numbers in Groq's history. What is now also gated is a THIRD-PARTY model (MiniMax M2.7) that is neither Meta's nor Groq's. So the pattern is not 'Llama is being retired'; it is that Groq's self-serve pricing surface is narrowing toward its own first-party-favouring and small open-weight rows (the gpt-oss family and Qwen) while larger and third-party models move behind a sales conversation. A model being gated is therefore NOT evidence that the model is being deprecated, and a model being absent from the public table is not evidence that it is unmaintained. WHAT A READER SHOULD DO WITH IT: never quote a Groq price for a Llama model from memory. The most commonly circulated Groq numbers in blog posts, cost comparisons and older revisions of this repo are Llama 3.3 70B and Llama 3.1 8B rates that were public before this gating and are no longer obtainable from the vendor at any published price. Treat any such figure as UNVERIFIABLE TODAY rather than as a stale-but-real rate. Note that Groq does still publish speed for gated rows (560, 280 and 260 tok/s above), so throughput is knowable while price is not - an unusual inversion, and a good argument for benchmarking rather than trusting recalled rates. ALSO RECORDED WHILE TRUE: the published tok/s figures are list attributes that can change with capacity, and the content-moderation rows are priced per 1M CHARACTERS ($22.00 English, $40.00 Arabic Saudi) rather than per 1M tokens, so tokens are the wrong unit for them entirely. No GPU-hour rate is published.

Notes

Two things make Groq's page structurally different from the others. First, its headline Llama models are NOT on the public list price at all: llama-3.1-8b-instant and llama-3.3-70b-versatile both read literally 'ContactSales' in the price column. So the cheapest commonly-quoted Groq number in circulation is stale - Llama 3.3 70B on Groq moved to sales-gated pricing, and only the gpt-oss / Qwen / guard models remain self-serve. Second, Groq publishes SPEED next to price on the same row (1000 and 500 tok/s for the gpt-oss models), which makes it the only provider here where throughput is a published list attribute rather than something you benchmark yourself. For character-priced models, tokens are the wrong unit - $22 per 1M CHARACTERS on a TTS-style model is not comparable to a $2 per 1M token input rate at all. No GPU hourly rate is published.

EXPORT CONTROL (2026-10-04): NOT APPLICABLE IN THE HARDWARE-SHIPMENT SENSE - this record prices a token-based API, so no accelerator is transferred across a border and the ECCN 3A090 hardware licence regime does not attach to the transaction. Recorded null rather than false precisely because a token API is where export control instead becomes an END-USER/TERMS question: the provider applies geographic and end-user restrictions to service access (the AI Diffusion Rule briefly put a tiered country structure on exactly this kind of service before being rescinded on 2025-05-13), and a customer can still be refused under sanctions or end-use policy regardless of where the compute sits. Anyone reading this record for a restricted destination should look at the provider's terms, not at the EAR. Sources for the regime: spl-bis-2025-05-ai-diffusion-rescission, spl-bis-2026-05-d5-guidance.

EXPORT CONTROL (2026-10-04): NOT APPLICABLE IN THE HARDWARE-SHIPMENT SENSE - this record prices a token-based API, so no accelerator crosses a border and the ECCN 3A090 hardware licence regime does not attach to the transaction. Recorded NULL rather than false precisely because a token API is where export control instead becomes an END-USER and TERMS question: the provider applies geographic and end-user restrictions to service access (the AI Diffusion Rule briefly put a tiered country structure on exactly this kind of service before being rescinded and unenforced from 2025-05-13), and a customer can still be refused under sanctions or end-use policy regardless of where the compute sits. For a restricted destination, read the provider's terms, not the EAR. Sources: spl-bis-2025-05-ai-diffusion-rescission, spl-bis-2026-05-d5-guidance.

LEAD TIME: null - correct and not a gap. A token-priced serverless API is provisioned on request, so there is no lead time to publish; the binding constraint when a model is at capacity is provider-side queuing and any per-account concurrency limit, which this repo records in notes where a vendor publishes one (see the DeepSeek concurrency figures already here). Deliberately NOT recorded: any 'model capacity lead time' estimate, because vendors publish none and the SEO 'GPU lead time' numbers that circulate do not apply to a rented endpoint. Probe log in source spl-negative-no-oem-publishes-lead-time.

ACCELERATOR LINK ADDED 2026-10-04: Groq serves on its own LPU ([[accelerators/groq-lpu]]), not on a rented GPU, which is why Groq is the one provider here where published throughput (1000 tok/s gpt-oss-20b, 500 tok/s gpt-oss-120b) is a list attribute rather than something you benchmark - there is no commodity GPU behind the rate to be throttled by.

VOLATILE STRUCTURE, RECORDED WHILE IT IS STILL TRUE - re-verified against the live model list 2026-10-04. THIS GATING PATTERN MOVES, and it is recorded here at a known date because it will not move in one direction. AS OF 2026-10-04, THREE rows on Groq's production and preview model tables read literally 'ContactSales' in BOTH the price column and the rate-limits column, and each is additionally tagged 'Enterprise' in the model list: (1) llama-3.1-8b-instant, 560 tok/s, 131,072 context; (2) llama-3.3-70b-versatile, 280 tok/s, 131,072 context, max completion 32,768; (3) minimaxai/minimax-m2.7, 260 tok/s, 196,608 context, max completion 131,072. Models 1 and 2 were already recorded; minimaxai/minimax-m2.7 was NOT in the previous revision of this record and is added by the 2026-10-04 re-verification. THE DIRECTION OF TRAVEL IS THE POINT, and it cuts against the intuitive reading. The FIRST Llama rows to move to ContactSales were Meta's own first-party open-weight models - the two most widely quoted numbers in Groq's history. What is now also gated is a THIRD-PARTY model (MiniMax M2.7) that is neither Meta's nor Groq's. So the pattern is not 'Llama is being retired'; it is that Groq's self-serve pricing surface is narrowing toward its own first-party-favouring and small open-weight rows (the gpt-oss family and Qwen) while larger and third-party models move behind a sales conversation. A model being gated is therefore NOT evidence that the model is being deprecated, and a model being absent from the public table is not evidence that it is unmaintained. WHAT A READER SHOULD DO WITH IT: never quote a Groq price for a Llama model from memory. The most commonly circulated Groq numbers in blog posts, cost comparisons and older revisions of this repo are Llama 3.3 70B and Llama 3.1 8B rates that were public before this gating and are no longer obtainable from the vendor at any published price. Treat any such figure as UNVERIFIABLE TODAY rather than as a stale-but-real rate. Note that Groq does still publish speed for gated rows (560, 280 and 260 tok/s above), so throughput is knowable while price is not - an unusual inversion, and a good argument for benchmarking rather than trusting recalled rates. ALSO RECORDED WHILE TRUE: the published tok/s figures are list attributes that can change with capacity, and the content-moderation rows are priced per 1M CHARACTERS ($22.00 English, $40.00 Arabic Saudi) rather than per 1M tokens, so tokens are the wrong unit for them entirely. No GPU-hour rate is published.

Cross-references

Accelerators (1)