GILD verdict
Choose on batch timing and retention, since both list gpt-oss-120b at the same price.
Fireworks lists a US$0.015 cached-input rate for gpt-oss-120b, which Groq's models page does not list, and fits open-model work that must keep no data by default outside the Response API. Groq fits batch jobs that can wait 24 hours to 7 days for 50% off without touching standard rate limits. At 100M input and 20M output tokens a month with no cache hits, both cost US$27 by GILD's calculation.
On gpt-oss-120b the two list the same US$0.15 input and US$0.60 output per 1M tokens, so 100M input and 20M output tokens a month cost US$27 on either, by GILD's calculation. Fireworks also lists a cached-input rate of US$0.015 per 1M tokens and has Zero Data Retention by default. Groq states that its batch runs at 50% lower cost within a 24-hour to 7-day window with no impact on standard rate limits, and that it retains no customer data for inference requests by default, though it retains data for batch jobs, fine-tuning and LoRAs, and when needed to protect platform reliability.
Choosing between Fireworks AI and Groq
Choose Fireworks AI if
- Your gpt-oss-120b prompts repeat, so the cached-input rate of US$0.015 per 1M tokens lowers the input bill.
- Prompts on open models should not be logged or stored by default, and on the Response API you will set store=False, since it keeps conversations up to 30 days by default.
- Offline jobs on models that support the Batch API can run at 50% off Serverless per-token prices.
Choose Groq if
- Batch work can wait 24 hours to 7 days for a 50% lower cost without touching your standard rate limits.
- Inference requests must keep no customer data by default, and your batch and fine-tuning use fits Groq's stated retention for those features.
Side by side
| Dimension | Fireworks AI | Groq |
|---|---|---|
| List price, gpt-oss-120bList prices in US dollars on November 1, 2026. | gpt-oss-120b Standard: US$0.15 input, US$0.015 cached input, US$0.60 output per 1M tokens1 | gpt-oss-120b: US$0.15 input, US$0.60 output per 1M tokens5 |
| Monthly cost for one workload (GILD's calculation from published rates)GILD's calculation from the list prices on Fireworks' and Groq's pages, November 1, 2026. Inputs: 100M input tokens with no cache hits and 20M output tokens a month on gpt-oss-120b, at US$0.15 input and US$0.60 output per 1M tokens on both. | At 100M uncached input and 20M output tokens a month on gpt-oss-120b: 100 x US$0.15 + 20 x US$0.60 = US$27 a month1 | The same workload: 100 x US$0.15 + 20 x US$0.60 = US$27 a month5 |
| Batch discount | Batch API at 50% off serverless per-token prices. Not all models support the Batch API2 | Batch API at 50% lower cost, with a 24-hour to 7-day processing window6 |
| Default data retention | Zero data retention by default for open models. The Response API keeps conversation data up to 30 days by default (store=True) unless store=False is set3 | Does not retain customer data for inference requests by default. Batch jobs, fine-tuning and LoRAs retain it, and Groq may retain data when needed to protect platform reliability, such as troubleshooting failures or investigating abuse7 |