Fireworks AI review: an inference provider built for production traffic
Answer in brief: Fireworks AI ranks #2 among the best AI inference providers. It is the strongest pick for production teams that want to tune speed against cost on every request. One endpoint offers Standard, Priority, Fast, Ultra-fast (preview), Batch (50% off) and Spot (30% off) modes. Reserved Throughput adds an SLA, and data retention is zero by default. gpt-oss-120B is $0.15 input / $0.015 cached / $0.60 output per 1M tokens.
Independent research, checked 2026-11-09 against Fireworks' pricing page and Forge announcements. We did not test the product hands-on.
Pick a model, pick a mode
Fireworks' standout idea is that one serverless endpoint serves several economic profiles. Every request bills on input, cached input and output tokens, and you switch modes per request:
- Standard: the default for most traffic.
- Priority: prioritized during peak load (gpt-oss-120B is $0.18 / $0.018 / $0.72).
- Fast: 100+ tokens per second for interactive apps.
- Ultra-fast: public preview since Forge on November 3, 2026, on NVIDIA Blackwell. Fireworks reports up to 1,500 tokens/sec on DeepSeek V4.1 Flash, and says a benchmarked 30-turn agent task dropped from one minute to eight seconds.
- Batch: 50% off input and output for work that can wait.
- Spot: 30% off Standard on best-effort capacity for select models.
Beyond serverless, Reserved Throughput guarantees capacity, billed in dollars per minute. On-demand deployments bill per GPU-second, and Fireworks Sovereign (private preview) runs Fireworks-managed inference on GPUs in your own cloud or data center.
Who Fireworks fits best
Fireworks fits product teams with steady production traffic, agent builders who care about cache hit rates, and enterprises that need compliance paperwork and Azure-friendly procurement.
Why teams pick Fireworks AI
- Automatic prompt caching with no write fee; cached input bills at 50% off or less. Fireworks says session affinity lets long agent runs reach cache hit rates of up to 99.6%.
- Reserved Throughput with a throughput SLA and 99.9% uptime terms. Overage bills at standard serverless rates, never a penalty rate.
- Security and compliance: zero data retention by default; SOC 2 Type II, ISO 27001, 27701 and 42001; HIPAA with a BAA.
- Easy procurement: Fireworks is a first-party provider in Microsoft Foundry, so usage counts toward your Azure consumption commitment. Volume discounts and post-paid invoicing are available.
- Spend control: monthly spend limits, Auto Reload, and request tagging by team and project for chargeback.
- Training on the same platform: Managed Training from $0.50 per 1M training tokens. You can serve the trained models on Serverless or on dedicated deployments.
The honest drawbacks
- On-demand GPU rates are high: H100 and H200 are $8.00/hr, B200 $13.00/hr, B300 $15.00/hr.
- Region-restricted deployments cost 1.5x, and US-only serverless carries a 50% premium on select models.
- Spot throughput can drop to zero, so you need a retry or fallback path.
- Ultra-fast and Sovereign are still in preview.
Pricing at a glance
| Item | Published price |
|---|---|
| gpt-oss-120B (Standard) | $0.15 input / $0.015 cached / $0.60 output per 1M |
| GLM 5.3 | $1.40 / $0.26 cached / $4.40 |
| Kimi K3 | $3.00 / $0.30 cached / $15.00 |
| Serverless entry price | From $0.05 per 1M input tokens |
| GPUs (on-demand) | H100 $8.00/hr, B200 $13.00/hr, GB300 $20.00/hr |
Best Fireworks AI alternatives
- Together AI: a broader catalog and cheaper dedicated H100s ($5.49/hr).
- Baseten: a lower gpt-oss-120b list price and strong custom-model deployments.
- Groq: low-latency self-serve inference at the same $0.15 / $0.60 gpt-oss-120b price.
Fireworks AI FAQ
How does Fireworks billing work?
You buy credits, and usage draws them down: per token on Serverless and per GPU-second on on-demand deployments. You can set a monthly spend limit.
Does Fireworks retain my data?
Fireworks states zero data retention by default and lists SOC 2 Type II, ISO 27001, 27701, 42001 and HIPAA with a BAA.
How fast is Fireworks Ultra-fast?
Fireworks reports up to 1,500 tokens/sec on DeepSeek V4.1 Flash in public preview. This is a vendor figure for a specific model and workload.
Sources
- Fireworks AI pricing
- Fireworks Forge announcements (November 3, 2026)
- Forge event page
- Fireworks trust center
← Back to the ranked list of the best AI inference providers
Named on this page as a provider? Join the network for a linked profile, Visit buttons and featured placement.