Best Inference Provider for gpt-oss-120b in Production 2026
The short answer
Fireworks for most production teams, Cerebras when latency is the product, Groq for speed at the standard price, Bedrock if you must stay inside AWS. gpt-oss-120b is OpenAI’s Apache 2.0 open-weight reasoning model (117B parameters, 5.1B active, released August 2025), and more than 20 hosts serve it. They agree on a price of about $0.15 input / $0.60 output per million tokens; they disagree on speed, caching, rate limits and how faithfully they run the model. All prices below are list USD read from vendor pages on October 16, 2026.
The comparison
| Provider | Input / cached / output (per 1M) | Measured speed (Artificial Analysis) | Context | Best for |
|---|---|---|---|---|
| Fireworks | $0.15 / $0.015 / $0.60 (Priority $0.18 / $0.018 / $0.72; batch −50%) | not in AA’s October snapshot | 131K | Default production pick |
| Cerebras | $0.35 / — / $0.75 | ~1,708 tok/s, 0.43 s to first token | 131K paid (65K free) | Real-time and voice agents |
| Groq | $0.15 / $0.075 / $0.60 | ~474 tok/s, 0.73 s | 131K, 65K max output | Fast at the standard price |
| Together AI | $0.15 / — / $0.60 | ~89 tok/s, 0.56 s | 131K | Teams already on Together |
| Amazon Bedrock | $0.1545 / — / $0.618 (Sydney Standard; Flex and Batch −50%) | ~88 tok/s | 128K | AWS IAM, VPC, contracts |
| Baseten | blended ~$0.07 (AA) | ~253 tok/s, 0.20 s | 128K | Lowest time to first token |
| DeepInfra / Novita / CoreWeave | blended $0.02–$0.04 (AA) | 36–228 tok/s | 131K | Batch jobs, lowest cost |
“Blended” is Artificial Analysis’s single price for a typical input-heavy mix; the vendor-page rows show separate input and output rates.
Provider notes
Fireworks. The most complete billing for production: a separate cached-input rate ($0.015, a 90% discount — the deepest of the major hosts), a Priority serving path for traffic that must not be throttled, and batch inference at 50% of serverless prices. For agents that resend long system prompts on every call, the cached rate often matters more than the headline price.
Cerebras. Its wafer-scale chips are the fastest measured route to gpt-oss-120b — roughly 3.5x Groq and nearly 20x the slowest GPU hosts. You pay for it: $0.35 / $0.75 is more than twice the standard input price. Its docs cap output at 40K tokens on paid plans, and the free tier is limited to 30 requests per minute and 1 million tokens a day, so it is not a free production option.
Groq. The standard $0.15 / $0.60 price with LPU speed (~474 tok/s measured), a 50% cached-input rate and the largest max-output allowance in this table (65,536 tokens). It is the best value when you want fast responses but do not need Cerebras-level throughput.
Amazon Bedrock. Priced at the standard rate, with Flex and Batch tiers at half price and Priority at a 75% premium. Choose it for procurement and security reasons — IAM, private networking, existing AWS commitments — not for speed.
Baseten. Artificial Analysis measured the lowest time to first token of any gpt-oss-120b host (0.20 s). Baseten also offers dedicated deployments, which matter once you need guaranteed capacity rather than shared serverless.
Budget hosts. CoreWeave, DeepInfra and Novita list blended prices of a few cents per million tokens. They suit offline batch work; check their speed and accuracy numbers before putting them behind a user-facing agent.
Check accuracy, not just price
The same weights do not guarantee the same answers. Artificial Analysis’s endpoint accuracy check put most hosts between 84% and 98% of the reference in its October 2026 snapshot, with outliers at 72% and 77%. Differences come from quantization, the harmony chat format gpt-oss requires, and tool-call parsing. Before you commit:
- Run your own eval set (tool calls included) against two or three hosts.
- Set
reasoning_effortexplicitly — hosts default to medium, and effort changes both cost and quality. - Keep an OpenAI-compatible client so you can fail over between hosts.
How to choose
- Production agent, cost-aware, long repeated prompts: Fireworks.
- Voice, live chat or anything where users watch tokens stream: Cerebras.
- Fast responses at the standard price: Groq.
- Enterprise on AWS: Bedrock.
- Overnight batch: a budget host, or Fireworks/Bedrock batch at 50% off.
For the wider provider picture see the cheapest inference providers for open-weight models and best LLM inference providers ranked; for closed-model rates see current API prices.
Last verified: October 16, 2026. Prices are list USD per million tokens from vendor pages; speed and accuracy figures are Artificial Analysis measurements read the same day.