Fireworks AI

Fireworks AI hosts open-weight models behind an OpenAI-style API, billed per token, and its docs state zero data retention by default for open models. On 2026-10-31 gpt-oss-120B costs $0.15 in and $0.60 out per million tokens, and DeepSeek V4.1 Flash costs $0.30 and $1.20. US-only serving costs 1.5 times more.

Facts at a glance

Prices checked 2026-10-31. Fireworks serves other labs' open models, so model quality comes from the maker and the data terms come from Fireworks.

Models and prices

Fireworks bills every text or vision request across input, cached input and output tokens. The table gives input, cached input and output per million tokens at Standard, with the Priority price alongside.

Fireworks also sells Fast variants of some models as adjacent rows, for example Kimi K3 Fast at $4.50 in and $22.50 out and GLM 5.3 Fast at $2.10 and $6.60, and it sells embeddings by parameter count: $0.10 per million tokens under 4B parameters, $0.20 from 4B to 16B, and $0.90 above 16B. A Rune 26B-A4B model is listed with a very low input price of $0.042. The page lists "headline models", so models not shown are priced on the model's own page.

US-only serving

For several models Fireworks lists a "(US)" variant, and the price is 1.5 times Standard. DeepSeek V4.1 Flash is $0.30 in and $1.20 out, while the US version is $0.45 and $1.80. GLM 5.3 is $1.40 and $4.40, and the US version is $2.10 and $6.60. Kimi K3 goes from $3.00 and $15.00 to $4.50 and $22.50. If a customer contract says data must be processed in the United States, this premium is the price of the clause. It is also a useful number to quote when you decide whether the clause is worth asking for. Fireworks has a separate US-only serverless page that we did not read, so confirm what the "(US)" models guarantee before you write it into a contract.

The three serving modes

Fireworks sells the same model at three speeds, and the pricing page points to a Serverless Modes page for selection details. Standard is the default. Priority costs 25 percent more on the headline models we compared, for example $0.30 becoming $0.375, and is the lever Fireworks names for reducing 503 overload errors. Fast is a premium variant for selected models. Batch inference is half price and asynchronous. Fine-tuned models cost the same to serve as the base model. The docs say Priority is prioritised above Standard traffic and less likely to be load shed, and that Fast variants aim for 100 or more tokens per second; neither page we read states a latency guarantee.

Rate limits and spend tiers

Fireworks rate limits work differently from most vendors, so read this section before you plan capacity.

Adaptive limits, per model. Limits are per account and per model, and there is no single account-wide token pool. Each model has its own tokens-per-minute limits, and each model's limit grows, shrinks and resets based only on your recent usage of that model. Using one model does not consume another's limit. Three measures apply, separately per model: total prompt tokens per minute (cached plus uncached), uncached prompt tokens per minute, and generated tokens per minute. The base ceilings depend on a model's parameter count, because smaller models get higher ceilings.

The docs say you can see current limits with firectl quota list. Fireworks describes an Elastic Burst feature that can lift the current limit above the ceiling for a time. Over a limit you get a 429. Overload returns a 503, and upgrading to Priority lowers the likelihood.

Spend tiers. Your tier sets base quotas and, for legacy postpaid accounts, the maximum monthly spend.

Prepaid accounts, which is the default billing, can set a monthly spend limit independent of tier. For guaranteed capacity there is Reserved Throughput, described as dedicated serverless throughput backed by an SLA. The docs say it is priced per token, traffic above the reservation runs pay as you go at list price with no SLA, and unused throughput does not roll over.

Data retention and training

This is where Fireworks is easiest to assess, because the docs are direct.

• Zero data retention by default. Fireworks says it does not log or store prompt or generation data for any open model without explicit user opt-in.
• How it works. Prompt and generation data exist only in volatile memory for the duration of the request. If prompt caching is active, some prompt data and the associated KV caches can sit in volatile memory for several minutes. Nothing is written to persistent storage.
• Metadata. Fireworks logs metadata, such as the token count of a request, as needed to run the service.
• Opt-in logging. Users can opt in to logging prompt and generation data for advanced features such as its FireOptimizer.
• Enterprise enforcement. Enterprise account admins can turn on a Zero Data Retention policy that rejects requests and jobs that would persist or log customer content.
• The Responses API is the exception. With store=True, the default, conversation data is kept: user prompts, model responses, and tool outputs. It is retained up to 30 days and deleted no later than that. Stored responses sit in a cache and can be evicted earlier, often within about an hour. Set store=False to opt out, or delete a record at once with the DELETE endpoint and the response_id.

The Responses API rule is the trap. A team that reads "zero retention" and then uses the stateful Responses API with its default has 30 days of stored conversations. Set store=False in your client wrapper so the safe value is the default. Fireworks says the Responses policy is designed to be consistent with OpenAI's API. The data handling page covers retention, and a statement on training belongs in your contract.

Regions

Deployments are multi-region by default, and Fireworks can move and spread them across regions. The multi-regions are GLOBAL, US, CANADA, EUROPE and APAC, which group single regions. For specific needs you can choose single regions such as US_IOWA_1 or EU_FRANKFURT_1, with listed accelerator types. The docs show H100, H200, B200, B300 and GB300 hardware across Asia-Pacific, Europe and North America, and a few regions with AMD hardware. Training is available in only some. Use a single region only when you must, because multi-region gives elastic scaling and resilience.

Living with adaptive limits

Most vendors give you a table of limits by tier. Fireworks gives you a moving ceiling per model, so capacity planning looks different. Four habits help.

1. Plan per model, not per account. Because each model has its own pool, a workload split across two models has two separate ceilings. Hitting a 429 on one model does not slow the other, which makes a second model a natural fallback even on the same host.
2. Watch uncached tokens. The uncached prompt limit is a quarter of the total prompt limit at every size. An agent that re-sends a long prefix with good cache hits can push far more total tokens than one that changes its prompt each turn. Keep the stable part of the prompt at the front and the changing part at the end.
3. Ramp up, do not spike. Adaptive means the limit grows with your recent usage of the model. A job that jumps from nothing to a large share of the ceiling is the one most likely to see 429s. Warm up with a smaller wave first.
4. Know the two errors. A 429 means you are over a limit, so slow down. A 503 means the platform is overloaded, and Priority tier is the documented way to reduce it. Retry both with exponential backoff and a cap.

The base ceilings are large. A small model, under 800 billion parameters, starts at 216 million total prompt tokens a minute, which few teams reach. The tight numbers are on the large models, where generated tokens per minute start at 216K. A very large model driving a long-output agent loop is where a ceiling shows up first.

Fine-tuning and training

Fireworks also sells training. Its pricing page lists supervised and preference fine-tuning, priced per million training tokens, and a serverless training option for LoRA, preference optimisation and reinforcement learning on a shared trainer pool. The page states that serving a fine-tuned model costs the same as the base model, which removes the usual surcharge for custom weights. Estimate training tokens as tokens in the dataset times epochs, and the docs tell you to multiply for the average number of conversation turns when you tune with reasoning traces, since multi-turn examples are unrolled. A promotional line on the page offers Muse Glimmer 30B SFT free through the UI and the serverless training API until 30 November, with no credit card needed.

Training sends your data to Fireworks for a longer time than inference does, and the zero-retention default is about inference. Check the training data terms before you upload customer data, and use an enterprise agreement if the data is sensitive.

Moving to Fireworks from another provider

1. Swap base URL, key and model id. The model id has the form accounts/fireworks/models/<name>, so check the exact string on the model page.
2. Choose the serving mode. Start on Standard, and move the user-facing path to Priority if you see 503s.
3. Pick standard or US variants according to your data clause, and compare the price at the 1.5 times premium.
4. Set the data defaults. Use store=False with the Responses API, and confirm the zero-retention behaviour for every feature you plan to use.
5. Re-test tool calls and structured output. Support varies by model, so run your schemas on the exact model.
6. Compare cached rates with your current host, because they differ more than headline input prices. DeepSeek V4.1 Flash is $0.006 cached here, and the same on Together AI, while DeepSeek's own off-peak cache rate is $0.003.
7. Set a monthly spend limit on the prepaid account before you give a key to an agent.

Moving away is the reverse. Because the interface follows the OpenAI shape, a switch to Groq or Together AI is mostly a base URL and model id change. The part that does not transfer is any custom model you fine-tuned on Fireworks, so keep the training data and the adapter files in your own storage.

Which agents can use it, and how

The API is OpenAI-style. The base URL is https://api.fireworks.ai/inference/v1.

curl https://api.fireworks.ai/inference/v1/chat/completions \
  -H "Authorization: Bearer $FIREWORKS_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"accounts/fireworks/models/gpt-oss-120b","messages":[{"role":"user","content":"Review this diff."}]}'

Any agent with a custom OpenAI-compatible provider option can use it: OpenCode, Cline and others. Fireworks also publishes a "Which model should I use?" page that covers picking an open model for chat, coding, vision, embeddings and agents, and migrating off Claude. Model IDs take the form accounts/fireworks/models/glm-5p3.

Alternatives

A worked cost example

Take 10 million input tokens and 2 million output tokens. The arithmetic is ours, from the Standard prices, with no caching.

Batch would halve each row. A cache hit on DeepSeek V4.1 Flash costs $0.006 per million, so a repeated 8 million token prefix is about five cents instead of $2.40. The US column is the 1.5 times premium applied to the Standard column.

Who should pick Fireworks

A pre-launch checklist

1. Set store=False on every Responses API call unless you want storage.
2. Decide on US-only serving before choosing model variants.
3. Check firectl quota list for your per-model TPM.
4. Use prepaid credits with a monthly spend limit.
5. Move latency-sensitive traffic to Priority.
6. Put a second model on another host behind a fallback.
7. Get the training and subprocessor statements in your contract.

Our take

Fireworks has the clearest data-retention page of the open-model hosts we read. Zero retention by default, a stated exception for the Responses API, and an enterprise switch that rejects anything that would persist. The price list is also unusually transparent, with Priority, Fast, batch and US variants shown. The costs are the 1.5 times US premium, per-model adaptive limits that take some learning, and a catalogue narrower than Together's. For a team that wants open models and a data page it can hand to legal, it is the one we would start with.

We have not run tests for this page. It reflects Fireworks' published documentation on 2026-10-31. Fireworks does not pay for placement here.

Related pages

Together AI, Groq, LLM API prices, open source LLM, private AI for teams.