inference/research
verifiedmedium · 0.80updated 2026-10-17

GroqCloud serverless inference

groq

Fields

id
services/svc-groqcloud
Provider
groq
Endpoint
https://api.groq.com/openai/v1
Input $/Mtok
0.15
Output $/Mtok
0.6
Price basis
USD per 1M TOKENS, representative model openai/gpt-oss-120b production list price, observed 2026-10-14 ($0.15 in / $0.60 out). gpt-oss-20b is $0.075/$0.30. The flagship Llamas are NOT self-serve: llama-3.3-70b-versatile and llama-3.1-8b-instant both read 'ContactSales' in the production table - the same sales gate first recorded in supply record rent-groq-api-token-pricing (2026-10-04) and still present today - yet Groq still serves llama-3.3-70b through OpenRouter at $0.59/$0.79, so the model is reachable, just not direct-billed. Speech models price per hour ($0.111/hr whisper-large-v3), not per token.
Context window
131,072
TTFT (ms)
1,000
Output tok/s
305.7
Measured by
artificial-analysis
Regions
us
notes
2026-10-17 fifth OpenRouter re-read, telemetry only. OpenRouter's Groq row for llama-3.3-70b-instruct is now 0.52s P50, 191 tok/s, 100.00% uptime, still $0.59/$0.79 (was 0.51s / 198 tok/s on the fourth read the same day). That is a routed price and a different workload from the Artificial Analysis median of 305.7 tok/s / 1.00s first chunk, re-confirmed unchanged this pass, which measured_* still carries. Groq-direct list price was not re-fetched; the price fields stay on openai/gpt-oss-120b at $0.15/$0.60. 2026-10-17 first-party reconfirmation. console.groq.com/docs/models still badges llama-3.3-70b-versatile Enterprise with Contact Sales, vendor speed 280 t/s, context 131072, max completion 32768. openai/gpt-oss-120b is still the public production rate at $0.15/$0.60 (vendor 500 t/s, 250K TPM / 1K RPM), gpt-oss-20b $0.075/$0.30 (vendor 1000 t/s), so the price fields do not move. groq.com/pricing no longer renders a rate card. OpenRouter's Groq row for Llama 3.3 70B the same day is $0.59/$0.79, 0.50s P50, 198 tok/s, 100% uptime — a routed price and a different workload from the Artificial Analysis median of 305.7 tok/s / 1.00s first chunk that measured_* still carries. 2026-10-17: Groq's llama-3.3-70b-versatile page is still badged Enterprise, context 131072, max output 32768, vendor speed about 280 tps, and still has no public per-1M price. The price fields stay on openai/gpt-oss-120b at $0.15/$0.60, the public production rate from 2026-10-14, because the Llama row is sales-gated. OpenRouter's Groq row for Llama 3.3 70B was $0.59/$0.79 on 2026-10-16, which is a routed price, not Groq-direct. measured_tps stays the Artificial Analysis median of 305.7 on Llama 3.3 70B, not the vendor 280. 2026-10-16 measurement refresh only. Artificial Analysis's Llama 3.3 70B provider page puts Groq first on output speed at 305.7 tok/s median, with 1.00s median time to first chunk — slower to first token than Vertex (0.68s) and CoreWeave (0.82s) on the same page, fastest once tokens flow. The previous figures (0.40s P50, 233 tok/s) were OpenRouter's own routing telemetry on 2026-10-14 and are a different workload, not a retracted number. Price was NOT re-verified: Groq's models page fetched the same day lists Llama 3.3 70B and Llama 3.1 8B as Enterprise with 'Contact Sales' in the price column, and prints ~500 tps as a vendor speed claim for GPT-OSS 120B (vendor, not used here). The $0.59/$0.79 figure therefore stays the 2026-10-14 OpenRouter Groq-row price, which is what a buyer routing through OpenRouter actually pays; Groq-direct self-serve pricing was not on the public page. Speed is Groq's reason to exist and it publishes the number itself: ~500 t/s on gpt-oss-120b, ~1000 t/s on gpt-oss-20b, 280 t/s on the (gated) llama-3.3-70b - vendor figures. The measured values recorded here are third-party: OpenRouter's 2026-10-14 P50 on Groq's actual endpoint, 0.40s latency / 233 tok/s / 100% 3d uptime, the fastest row on the llama-3.3-70b table; the same figure corroborates via Artificial Analysis (~233 tps median vs the 280 t/s marketing spec). Rate limits on the developer tier are published per model (250K TPM / 1K RPM for gpt-oss). LPX pairs Groq LPUs with NVIDIA GPUs; the endpoint does not expose which serves a given request.

Notes

2026-10-17 fifth OpenRouter re-read, telemetry only. OpenRouter's Groq row for llama-3.3-70b-instruct is now 0.52s P50, 191 tok/s, 100.00% uptime, still $0.59/$0.79 (was 0.51s / 198 tok/s on the fourth read the same day). That is a routed price and a different workload from the Artificial Analysis median of 305.7 tok/s / 1.00s first chunk, re-confirmed unchanged this pass, which measured_* still carries. Groq-direct list price was not re-fetched; the price fields stay on openai/gpt-oss-120b at $0.15/$0.60. 2026-10-17 first-party reconfirmation. console.groq.com/docs/models still badges llama-3.3-70b-versatile Enterprise with Contact Sales, vendor speed 280 t/s, context 131072, max completion 32768. openai/gpt-oss-120b is still the public production rate at $0.15/$0.60 (vendor 500 t/s, 250K TPM / 1K RPM), gpt-oss-20b $0.075/$0.30 (vendor 1000 t/s), so the price fields do not move. groq.com/pricing no longer renders a rate card. OpenRouter's Groq row for Llama 3.3 70B the same day is $0.59/$0.79, 0.50s P50, 198 tok/s, 100% uptime — a routed price and a different workload from the Artificial Analysis median of 305.7 tok/s / 1.00s first chunk that measured_* still carries. 2026-10-17: Groq's llama-3.3-70b-versatile page is still badged Enterprise, context 131072, max output 32768, vendor speed about 280 tps, and still has no public per-1M price. The price fields stay on openai/gpt-oss-120b at $0.15/$0.60, the public production rate from 2026-10-14, because the Llama row is sales-gated. OpenRouter's Groq row for Llama 3.3 70B was $0.59/$0.79 on 2026-10-16, which is a routed price, not Groq-direct. measured_tps stays the Artificial Analysis median of 305.7 on Llama 3.3 70B, not the vendor 280. 2026-10-16 measurement refresh only. Artificial Analysis's Llama 3.3 70B provider page puts Groq first on output speed at 305.7 tok/s median, with 1.00s median time to first chunk — slower to first token than Vertex (0.68s) and CoreWeave (0.82s) on the same page, fastest once tokens flow. The previous figures (0.40s P50, 233 tok/s) were OpenRouter's own routing telemetry on 2026-10-14 and are a different workload, not a retracted number. Price was NOT re-verified: Groq's models page fetched the same day lists Llama 3.3 70B and Llama 3.1 8B as Enterprise with 'Contact Sales' in the price column, and prints ~500 tps as a vendor speed claim for GPT-OSS 120B (vendor, not used here). The $0.59/$0.79 figure therefore stays the 2026-10-14 OpenRouter Groq-row price, which is what a buyer routing through OpenRouter actually pays; Groq-direct self-serve pricing was not on the public page. Speed is Groq's reason to exist and it publishes the number itself: ~500 t/s on gpt-oss-120b, ~1000 t/s on gpt-oss-20b, 280 t/s on the (gated) llama-3.3-70b - vendor figures. The measured values recorded here are third-party: OpenRouter's 2026-10-14 P50 on Groq's actual endpoint, 0.40s latency / 233 tok/s / 100% 3d uptime, the fastest row on the llama-3.3-70b table; the same figure corroborates via Artificial Analysis (~233 tps median vs the 280 t/s marketing spec). Rate limits on the developer tier are published per model (250K TPM / 1K RPM for gpt-oss). LPX pairs Groq LPUs with NVIDIA GPUs; the endpoint does not expose which serves a given request.

Cross-references

Models (4)