Skip to content
ReviewRamble

The weekly brief

AI, reviewed, in your inbox.

  • New reviews of tools, models, MCP servers and repos, with the verdict up front
  • Re-checked prices and ratings when something changes
  • New analysis pieces as they are published

Models, agents, MCP servers and the tools actually worth your time.

One email a week. Unsubscribe with the link in any brief. How we use your email

Fireworks AI

Fast hosted inference and fine-tuning for open models like Kimi, DeepSeek, GLM and GPT OSS, from shared API to dedicated GPUs.

Reviewed by ReviewRamble editors · Updated

Visit Fireworks AI Pay per use

Share this page:XLinkedInFacebookRedditWhatsAppEmail

Our rating

4.4/5

Fireworks is a safe production home for open models, keeping nothing by default and serving fine-tuned models at base prices, and its dedicated GPUs are the one expensive part.

What is Fireworks AI?

Fireworks AI runs open-weight AI models for developers. You can call popular models through a shared pay-per-token API, rent dedicated GPUs for your own deployment, or fine-tune a model and serve it at the same price as the original. Since November 2026 it also sells Nexus, which splits the work of coding agents such as Claude Code, Codex and OpenCode between its open models and the closed-model account you already have.

Prices are per million tokens. Examples from the docs: GPT OSS 120B costs $0.15 input and $0.60 output, DeepSeek V4.1 Flash $0.30 and $1.20, and Kimi K3 $3.00 and $15.00. Batch jobs cost half the normal price, and cached input is discounted by roughly 80% to 98% depending on the model. Dedicated H100 and H200 GPUs cost $8 an hour, billed by the second.

How much does Fireworks AI cost?

PlanPriceWhat you get
GPT OSS 120B$0.15 in / $0.60 out per 1M tokens
Kimi K3$3 in / $15 out per 1M tokens
Dedicated H100/H200$8/hour
Dedicated B200$13/hour
Fine-tuningFrom $0.50 per 1M training tokens
Nexus Pro$8/month
Nexus Teams$10/seat/month

Prepaid credits. Per 1M tokens, e.g. GPT OSS 120B $0.15 in / $0.60 out, Kimi K3 $3 / $15. Batch 50% off. Dedicated H100/H200 $8/hour, B200 $13/hour. Fine-tuning from $0.50 per 1M training tokens. Nexus plans from $8/month (Pro) or $10/seat/month (Teams).

Is Fireworks AI worth it?

Once you have decided to run an open model, the hard part is finding somewhere to serve it that keeps up with new releases and does not keep your data.

Fireworks does that job. The October 2026 catalogue tracks the main releases closely, with Kimi K3, DeepSeek V4.1 Flash, GLM 5.3, Qwen 3.8 Max, MiniMax M3 and NVIDIA Nemotron models. The API follows the OpenAI format, so moving an existing app is mostly a change of base URL and key.

Its data defaults are better than most rivals'. The docs say prompts and outputs for open models are not logged or stored unless you opt in; they live only in memory for the request, plus a few minutes of prompt cache. Fireworks lists SOC 2 Type II, HIPAA, and ISO 27001, 27701 and 42001, with reports in its trust centre.

The prices are detailed rather than hidden. Each model has a Standard rate, a Priority tier at roughly 1.25x for steadier response times, a Fast tier at 1.5x on some models, and a 50% premium for US-only routing. Batch is half price, and cached input falls to as little as $0.006 per million on DeepSeek V4.1 Flash, which changes the sums for agents that resend long contexts.

Fine-tuning, teaching a model on your own examples, is the strongest argument. LoRA training starts at $0.50 per million tokens for models up to 16B parameters, and the tuned model is then served at the base model's per-token price rather than on a dedicated GPU.

Fireworks now also reaches into the coding tools you already pay for. Nexus, expanded at its Forge event on 3 November 2026, plugs into Claude Code, Codex, OpenCode, Cursor and others, and its Auto mode scores each turn: routine ones go to a Fireworks open model, harder ones pass to your closed provider on your own key. Nexus Desktop, a Mac app launched the same day, replaces the terminal setup with one switch per tool, and a dashboard tracks spend by model, user and key against an all-Opus 5.5 baseline. Plans start at $8 a month for Pro and $10 a seat for Teams. Two caveats: Fireworks says the savings figures are estimates, not invoice amounts, and Claude Desktop support is limited.

The catch: two things cost more than you expect. The Responses API stores conversations for 30 days because store=True is the default, which catches people out. And dedicated hardware is where Fireworks loses: $8 an hour for an H100 against fal's $4.50 list rate, with the account-wide request ceiling fixed at 6,000 per minute whatever you spend.

Our take: Fireworks is like a well-run serviced flat. The shared rooms are tidy and fairly priced, and renting a whole floor to yourself costs a good deal more than it would down the road.

Where it earns its place

  • A backend engineer at a SaaS company with an OpenAI-based support bot

    Cut inference cost by moving to an open model

    Point the OpenAI SDK at Fireworks and switch to GPT OSS 120B at $0.15 in and $0.60 out per million tokens. Ten billion input and two billion output tokens a month is about $2,700 before cache discounts.

  • An ML team at a health-tech startup

    Fine-tune a model on de-identified clinical notes and serve it under HIPAA

    LoRA-tune a 16B-or-smaller model at $0.50 per million training tokens, then serve it at the base price. Zero retention is the default for inference; enterprise admins can also block training uploads entirely.

  • A data team at an e-commerce marketplace

    Classify and tag millions of product listings overnight

    Run batch jobs at 50% of serverless prices; on DeepSeek V4.1 Flash that is $0.15 in and $0.60 out per million tokens, with repeated system prompts cached at a fraction of a cent.

Before you commit

Setup
Add prepaid credits, create an API key and call any serverless model with the OpenAI SDK; no payment method limits you to 10 requests a minute.
Integrations
OpenAI-compatible chat, completions and Responses APIs, embeddings, batch, managed and serverless training APIs, and on-demand GPU deployments. Nexus connects Claude Code, Codex, OpenCode, Cursor, VS Code and Copilot, with a Mac desktop app.
Data and privacy
No logging of prompts or outputs for open models without opt-in; the Responses API stores data 30 days by default. SOC 2 Type II, HIPAA, ISO 27001, 27701 and 42001.
Cost at scale
Spending tiers run from $50 to $5,000 in credits to unlock higher limits; one dedicated H100 running all month is about $5,800, and region-restricted deployments cost 1.5x.
Cancelling and refunds
Serverless use is pay-as-you-go on prepaid credits; any subscription renews automatically for the same term until you cancel at least 3 days before the next term, in the app or by emailing support. Fees paid are non-refundable unless the law requires otherwise, including unused time and prepayments; what happens to your data after closing is not published.
Limits
Serverless limits are adaptive tokens-per-minute ceilings set by tier; the account-wide cap is 6,000 requests a minute, and everything pauses at 100% of your monthly spend limit.

What’s good

  • No prompt or output logging by default for open models
  • SOC 2 Type II, HIPAA, ISO 27001, 27701 and 42001
  • Fine-tuned models served at base-model token prices
  • Batch at 50% off; cached input up to 98% off

What’s not

  • Dedicated H100 at $8/hour, well above specialist GPU hosts
  • Responses API keeps conversations 30 days unless store=False
  • No free credits for new accounts on the pricing pages
  • Fixed 6,000 requests per minute account-wide ceiling

Best for

Teams running open models in production who want fine-tuning and predictable per-token pricing.

Fireworks AI alternatives

Fireworks AI: common questions

Is Fireworks AI worth it?
Fireworks is a safe production home for open models, keeping nothing by default and serving fine-tuned models at base prices, and its dedicated GPUs are the one expensive part. We rate it 4.4 out of 5.
How much does Fireworks AI cost?
Prepaid credits. Per 1M tokens, e.g. GPT OSS 120B $0.15 in / $0.60 out, Kimi K3 $3 / $15. Batch 50% off. Dedicated H100/H200 $8/hour, B200 $13/hour. Fine-tuning from $0.50 per 1M training tokens. Nexus plans from $8/month (Pro) or $10/seat/month (Teams).
What are the downsides of Fireworks AI?
Dedicated H100 at $8/hour, well above specialist GPU hosts; responses API keeps conversations 30 days unless store=False; no free credits for new accounts on the pricing pages; fixed 6,000 requests per minute account-wide ceiling.
Who is Fireworks AI best for?
Teams running open models in production who want fine-tuning and predictable per-token pricing.
How much does Fireworks AI cost for a team or in production?
Spending tiers run from $50 to $5,000 in credits to unlock higher limits; one dedicated H100 running all month is about $5,800, and region-restricted deployments cost 1.5x.
Is Fireworks AI safe for company data?
No logging of prompts or outputs for open models without opt-in; the Responses API stores data 30 days by default. SOC 2 Type II, HIPAA, ISO 27001, 27701 and 42001.
How do I cancel Fireworks AI, and can I get a refund?
Serverless use is pay-as-you-go on prepaid credits; any subscription renews automatically for the same term until you cancel at least 3 days before the next term, in the app or by emailing support. Fees paid are non-refundable unless the law requires otherwise, including unused time and prepayments; what happens to your data after closing is not published.
What are the best Fireworks AI alternatives?
The closest alternatives we have reviewed are Together AI, Groq, Cerebras and OpenRouter.

Answers checked Nov 18, 2026.

Make Fireworks AI? Show this rating on your siteFireworks AI rated 4.4 out of 5 on ReviewRamble