Fireworks AI review: what it is good at, and what it is not

What Fireworks actually sells

Three products, and they are priced on completely different units, which is the first thing to get straight.

The serverless tier is the one that matters for most readers. It speaks the OpenAI Chat Completions shape and the Anthropic Messages shape, which means a coding agent, an SDK, or a gateway configured for either vendor usually needs nothing more than a new base URL and key. That compatibility is table stakes now, but Fireworks does it cleanly.

The model catalog, and the hole in it

Fireworks serves the current generation of open and open-weight models: DeepSeek V4 in Pro and Flash variants, the Kimi K2.6, K2.7 Code and K3 line, GLM 5.1 and 5.2, Qwen 3.7 Plus and 3.8 Max, MiniMax M2.7 and M3, the gpt-oss 20B and 120B pair, and NVIDIA Nemotron models. Vision, image (FLUX.1), and audio (Whisper V3) round it out.

We keep a card per model with the independently published coding scores next to the list price, so DeepSeek V4 Pro and GLM 5.2 can be compared on quality before you decide which one Fireworks should serve you. Two structural things about the catalog itself are worth knowing before you build on it.

• Unlisted models are priced by size, not by name. Anything Fireworks hosts that is not on the individual rate card falls into a bracket: under 4B parameters at $0.10 per million, 4B to 16B at $0.20, above 16B at $0.90, and mixture-of-experts variants between $0.50 and $1.20. That is unusually transparent and it means a niche model has a knowable price before you deploy it.
• There are no first-party frontier models. You cannot buy Claude Opus 5 or gpt-5.6-sol from Fireworks the way you buy them from Anthropic and OpenAI. If your application needs a frontier model for the hard 20 percent of calls, Fireworks is one of two providers you will hold keys for, not the only one.

That second point is the honest limit of the whole open-model serving tier, not a Fireworks-specific failing. It is also why a gateway sits in front of most production setups. See what an LLM gateway does if you are about to hand-roll that routing.

Speed: the claims and the evidence

Fireworks markets on latency and publishes customer numbers rather than synthetic benchmarks. Notion reports moving from roughly 2 seconds to 350 milliseconds after fine-tuning on the platform, and Quora reports a 3x response-time speedup after migrating. Named users include Cursor, Vercel, Notion, Sourcegraph, UiPath, Quora, Genspark, and Cresta.

The tier structure is more informative than the marketing. Fireworks charging a 25 to 50 percent premium for Priority and Fast is an admission that Standard capacity is contended, which is true of every serverless provider and rarely stated. If your workload is latency sensitive, budget for Priority from the start rather than discovering Standard queueing under load.

Developer experience

Pointing an existing client at Fireworks is a base-URL swap. The example below is the shape most coding agents and SDKs want.

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/<model-id>",
    "messages": [{"role": "user", "content": "Explain this stack trace."}]
  }'

export OPENAI_BASE_URL="https://api.fireworks.ai/inference/v1"
export OPENAI_API_KEY="$FIREWORKS_API_KEY"

Model identifiers are namespaced (accounts/fireworks/models/<name>), which is more verbose than the bare slugs most providers use and occasionally trips up clients that validate model names against a hardcoded list. Copy the exact id from the model page rather than guessing it, and check your gateway or agent handles a slash-heavy model id before you commit.

Managed training is the other developer-facing surface worth knowing about: it bills per million training tokens rather than per GPU hour, and it deploys the resulting checkpoint automatically. LoRA supervised fine-tuning on a model up to 16B is $0.50 per million training tokens; DPO is double that; full-parameter runs are double again. For a small adapter that is a few dollars, which is a very different proposition from renting a GPU for a day.

Fireworks against Together and Baseten

These three are the closest comparison set, and on list price they are nearly indistinguishable. The differences that remain are structural rather than numeric.

Read the cached-input row rather than the base row if you run agents. Fireworks and Baseten both read GLM 5.2 cached input at $0.14, a tenth of base; Together charges $0.26, nearly a fifth. On a workload where 80 percent of input is cached repo context, that single line is worth more than any other difference on the table.

Together wins the DeepSeek V4 Pro comparison on input and loses it on output, so which of the three is cheapest genuinely depends on your token ratio. There is no universally cheapest option among them, which is a more useful conclusion than a ranking.

Where Fireworks is the wrong choice

• You want the cheapest tokens. Fireworks is at market rate, not below it. DeepInfra was cheaper on every comparable model we checked in August 2026, including $0.09 / $0.18 against Fireworks' $0.14 / $0.28 on DeepSeek V4 Flash.
• You want to rent raw GPUs. At $7.00 per H100 hour (rising to $8.00 on 1 September 2026), Fireworks on-demand is roughly double Modal's effective rate and triple DeepInfra's $2.20. Fireworks GPUs are priced as a convenience attached to the platform, not as a compute market.
• You need a frontier model as your primary. Go to the lab, or to an aggregator that can reach both.
• You want a generous evaluation budget. $1 in credits runs out during your first serious test.

Questions people ask

Keep reading

Continue through Inference providers with Fireworks AI pricing: every rate, tier, and free credit.