Fireworks.ai
Verified Oct 15, 2026 ActiveTrains and serves open and custom language models with per-token API pricing.
Listed on EffectHub ·
Pricing
- Pricing model
- Paid
Prepaid, usage-based credits with no subscription: serverless models not priced individually cost $0.10 per 1M tokens under 4B parameters, $0.20 for 4B-16B and $0.90 above 16B, while headline models carry their own input/cached/output rates (for example GPT OSS 120B at $0.15 / $0.015 / $0.60). Batch runs at 50% of serverless; on-demand GPUs start at $8 per hour for an H100 and fine-tuning from $0.50 per 1M training tokens.
Use tool ↗About Fireworks.ai
Fireworks provides infrastructure for training and running specialized language models: guided fine-tuning through custom reinforcement-learning loops, inference optimized for both open-source and proprietary models, and a library of pre-optimized models ready to call. It targets developers and enterprises building products like code assistants, conversational AI, agentic systems and multimodal applications, with customers named on the site including Cursor, Vercel, Notion and Sourcegraph. Access is through an OpenAI- and Anthropic-compatible API, with serverless pay-per-token pricing, on-demand dedicated deployments, and reserved capacity for guaranteed throughput. It was founded in 2022 by Lin Qiao, formerly head of PyTorch at Meta, along with several other ex-Meta engineers, and is based in the United States.
- Platforms
- Web · API
- API
- Yes
- Last verified
- Oct 15, 2026
Verified sources
- Serverless pricing docs: size-based rates of $0.10, $0.20 and $0.90 per 1M tokens (under 4B, 4B-16B, over 16B), per-model rates for headline models such as GPT OSS 120B $0.15 input / $0.60 output, batch at 50% (15 October 2026)docs.fireworks.ai ↗
- Pricing page: prepaid usage-based credits, on-demand H100 and H200 at $8.00 per hour, B200 $13.00; LoRA SFT fine-tuning $0.50 per 1M training tokens up to 16B; no free-credit offer shown (15 October 2026)fireworks.ai ↗
- Footer reads '© 2026 Fireworks AI, Inc.' (15 October 2026)fireworks.ai ↗
- Lin Qiao (ex-Meta PyTorch lead) and team founded Fireworks AI in 2022research.contrary.com ↗
Is Fireworks.ai legit?
We don’t label tools safe or unsafe. Below are the records we checked ourselves, each with the date we read it.
- Domain registered
- Mar 11, 2020 (6 years ago)Registry record (RDAP), registrar NameCheap, Inc., looked up Sep 17, 2026
- TLS certificate
- Issued by Let's Encrypt, valid until Jan 9, 2027Seen in our own connection to the site on Oct 15, 2026
- Last checked
- Oct 15, 2026The site answered
- On the market since
- 2022The vendor's own material, read by us
- Operated by
- Fireworks AI, Inc., United StatesOur research, card verified Oct 15, 2026
Pros & cons of Fireworks.ai
Fireworks is a credible, well-funded infrastructure player founded by the former head of Meta's PyTorch team, with real production customers like Cursor and Notion rather than just logo placement. Per-token pricing that scales with model size keeps small-model inference cheap, but there's no lasting free tier beyond a small starter credit, so this is squarely a pay-as-you-go developer tool rather than something to try casually.
Pros
- OpenAI- and Anthropic-compatible API, easing migration from those providers
- Pricing scales down for smaller models, from about $0.10 per million tokens
- Named production customers including Cursor, Vercel, Notion and Sourcegraph
- Founded by experienced ex-Meta PyTorch engineering leadership
- Reserved and on-demand deployment options alongside serverless
Cons
- No lasting free tier, only a small one-time starter credit
- Per-token pricing makes total monthly cost hard to predict without usage data
- Model-size-tiered pricing requires understanding which tier a chosen model falls into
Best for
- Developers building products that need low-latency inference on open or custom models
- Teams that want to fine-tune a model via guided workflows or custom RL loops
- Companies wanting an OpenAI/Anthropic-compatible API to swap in alternative models
Not for
- Hobbyists wanting to experiment for free beyond a small starter credit
- Teams wanting a flat, predictable monthly bill instead of usage-based token pricing
Watch out: The $1 starter credit is enough for roughly 1 million tokens on a large model, which is easy to burn through quickly once testing moves beyond a few requests.