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

Fireworks AI serverless inference

fireworks

Fields

id
services/svc-fireworks-serverless
Provider
fireworks
Endpoint
https://api.fireworks.ai/inference/v1
Input $/Mtok
0.15
Output $/Mtok
0.6
Price basis
USD per 1M TOKENS, representative model accounts/fireworks/models/gpt-oss-120b Standard tier, observed 2026-10-14 ($0.15 input / $0.015 cached input / $0.60 output). Fireworks prices THREE modes: Standard, Priority (+25% on every cell), and Fast (a separate, higher-priced serving path - Kimi K3 Fast is $4.50/$22.50 vs Standard $3.00/$15.00). Unlisted models fall back to size brackets: <4B $0.10, 4-16B $0.20, >16B $0.90, MoE <=56B $0.50, MoE 56.1-176B $1.20 (one number for input+output). Batch inference = 50% of serverless. '(US)' region-pinned variants of the headline models cost 1.5x the global price.
Context window
131,072
Regions
useu
notes
2026-10-17 re-read of docs.fireworks.ai/serverless/pricing: gpt-oss-120b Standard is still $0.15 / $0.015 cached / $0.60 (Priority $0.18 / $0.018 / $0.72). No Llama 3.3 row on the named table, so the representative model stays gpt-oss-120b and the price fields do not move. Size-band fallback for unlisted models is unchanged. No measured_ttft/tps recorded: Fireworks does not publish per-model speed on the pricing page and this pass did not land a third-party median for this provider - Artificial Analysis benchmarks Fireworks on Llama 3.3 70B but the live cells were not extractable today, so the fields stay null rather than carrying a second-hand number. 'Fast' serverless mode is the paid-latency tier; Standard is the default the basis prices. Fireworks also sells on-demand GPU deployments ($/hr) which are supply-type records, not this one.

Notes

2026-10-17 re-read of docs.fireworks.ai/serverless/pricing: gpt-oss-120b Standard is still $0.15 / $0.015 cached / $0.60 (Priority $0.18 / $0.018 / $0.72). No Llama 3.3 row on the named table, so the representative model stays gpt-oss-120b and the price fields do not move. Size-band fallback for unlisted models is unchanged. No measured_ttft/tps recorded: Fireworks does not publish per-model speed on the pricing page and this pass did not land a third-party median for this provider - Artificial Analysis benchmarks Fireworks on Llama 3.3 70B but the live cells were not extractable today, so the fields stay null rather than carrying a second-hand number. 'Fast' serverless mode is the paid-latency tier; Standard is the default the basis prices. Fireworks also sells on-demand GPU deployments ($/hr) which are supply-type records, not this one.

Cross-references

Models (9)