Skip to main content
Watch: I Built an AI App Without Leaving Databricks

Cost levers - verified

How to cut your Fireworks AI API bill

Prompt caching: automatic. Cache lifetime: several minutes to several hours. Batch: 50% off, 12, 24, 48 or 72 hours. Batch and caching stack: yes. Service tiers: Priority (1.2x to 1.5x), Fast (1.5x).

Read from docs.fireworks.ai. Compare every provider's cost levers

Caching
Automatic
Cache hit saves
See docs
Batch
50% off
Service tiers
2

Prompt caching

Mode

Automatic: repeated prefixes are cached with no code change

Enabled by default for all models and deployments; session affinity via the user field or x-session-affinity header; x-prompt-cache-isolation-key to isolate.

Cache lifetime

Usually at least several minutes, up to several hours depending on model, load and deployment; oldest evicted first

Minimum prompt
Not stated
Cache write cost
Not stated
Cache read cost

Serverless: default 50% discount on cached prompt tokens, varies by model. Dedicated: cached tokens close to free (capacity benefit).

Storage cost
Not stated
How to enable

Automatic; send a stable user ID or x-session-affinity header so requests land on the same replica.

Request fragment
x-session-affinity: <session id>
From the docs
“Prompt caching is enabled by default for all Fireworks models and deployments.”
docs.fireworks.ai
“For serverless models, cached prompt tokens are discounted compared to regular prompt tokens. The default discount is 50%, but the exact discount varies by model.”
docs.fireworks.ai
“Cached prompts usually stay in the cache for at least several minutes. Depending on the model, load level, and deployment configuration, it can be up to several hours.”
docs.fireworks.ai
“You may place this identifier either in the `user` field of the request body or in the `x-session-affinity` header.”
docs.fireworks.ai

Batch API

Discount

50% off

50% off serverless per-token prices (input and output); stacks with an additional 50% on cached tokens

Window

Job expiration selectable: 12, 24, 48 or 72 hours maximum (completed rows billed and saved on expiry)

Limits

Input dataset up to 80 GiB; output dataset max 8GB; per-request limits same as Chat Completions; models that support On-Demand Deployments

Endpoint

Batch Inference Job API (create-batch-inference-job), firectl, or UI

Stacks with caching

Yes: the batch discount and the cache discount combine

“Batch jobs automatically use prompt caching for additional 50% cost savings on cached tokens.”

Sources: docs.fireworks.ai/guides/batch-inference, docs.fireworks.ai/serverless/pricing

From the docs
“Process large volumes of requests asynchronously at 50% off Serverless per-token prices.”
“Batch inference is billed at 50% of serverless pricing on both input and output.”
“Batch jobs automatically use prompt caching for additional 50% cost savings on cached tokens.”
“Job expiration: Select from 12, 24, 48, 72 hours maximum in Optional Settings”
“Input dataset: Up to 80 GiB”
“Output dataset: Max 8GB (job may expire early if limit is reached)”

Sources: docs.fireworks.ai/guides/batch-inference, docs.fireworks.ai/serverless/pricing

Service tiers

Flex trades speed or availability for a lower price; Priority and Fast cost more for lower latency or reserved capacity.

Priority

Priority tier

Set with service_tier: priority

Price effect

Higher price, per model on the pricing table (headline models 1.2x to 1.5x Standard; most 1.25x)

Derived from the pricing table. Computed from the pricing page Standard vs Priority table: 1.25x for most rows (Ember-1, Kimi K3, DeepSeek V4.1 Flash, GLM 5.3, Nemotron), 1.5x for Qwen 3.8 Max and MiniMax M3, 1.2x for OpenAI GPT OSS 120B.

Trade-off

Higher reliability at peak; prioritized above Standard and less likely to be load shed (503); select models

From the docs
“Priority tier is for workloads that require higher reliability during peak traffic periods, at a higher price point. Priority tier is prioritized above Standard traffic and is less likely to be load shed (503 server overloaded).”
“To use priority tier, set `service_tier` to `"priority"`.”

Sources: docs.fireworks.ai/serverless/serverless-modes, docs.fireworks.ai/serverless/pricing

Fast

Fast

Set with Separate router model id (e.g. accounts/fireworks/routers/kimi-k3-fast)

Price effect

Higher price (pricing table: Kimi K3 Fast and GLM 5.3 Fast at 1.5x Standard)

Derived from the pricing table. Computed from the pricing table: Kimi K3 Fast $4.50/$0.45/$22.50 vs Kimi K3 $3.00/$0.30/$15.00; GLM 5.3 Fast $2.10/$0.39/$6.60 vs GLM 5.3 $1.40/$0.26/$4.40, both 1.5x.

Trade-off

Aims for 100+ tokens per second; same model and quality

From the docs
“Fast is a high-speed mode, useful for interactive applications that require fast response speeds, at a higher price point. Fast variants aim for 100+ tokens per second of generated throughput.”
“To use Fast, change the `model` ID as listed below.”

Sources: docs.fireworks.ai/serverless/serverless-modes, docs.fireworks.ai/serverless/pricing

Other discounts and surcharges

  • On-demand (dedicated) deployments

    “For dedicated deployments, prompt caching frees up resources, leading to higher throughput on the same hardware.”
    docs.fireworks.ai

Fireworks AI per-model prices

This site does not track per-model Fireworks AI prices. Use the sources below for the current rate card.

Research notes

Batch page also says 'Jobs expire after 24 hours' under Handling expired jobs, which conflicts with the selectable 12/24/48/72 hour expiration (24h is likely the default). Priority and Fast multipliers derived from the pricing table, not stated as rules.

FAQ

Does Fireworks AI cache prompts automatically?

Automatic: repeated prefixes are cached with no code change. Enabled by default for all models and deployments; session affinity via the user field or x-session-affinity header; x-prompt-cache-isolation-key to isolate. How to enable: Automatic; send a stable user ID or x-session-affinity header so requests land on the same replica.

How much does a cache hit cost on Fireworks AI?

Serverless: default 50% discount on cached prompt tokens, varies by model. Dedicated: cached tokens close to free (capacity benefit).. This site has no per-model Fireworks AI prices to compute it from.

Does Fireworks AI have a batch API discount?

50% off serverless per-token prices (input and output); stacks with an additional 50% on cached tokens. Window: Job expiration selectable: 12, 24, 48 or 72 hours maximum (completed rows billed and saved on expiry). Stacks with caching: yes: the batch discount and the cache discount combine.

What service tiers does Fireworks AI offer?

Priority: Higher price, per model on the pricing table (headline models 1.2x to 1.5x Standard; most 1.25x) (derived from the pricing table). Fast: Higher price (pricing table: Kimi K3 Fast and GLM 5.3 Fast at 1.5x Standard) (derived from the pricing table).

Sources

Last verified from Fireworks AI's own pages below. Quotes are verbatim; each was matched against the live page.

Newsletter

Know when a provider changes its prices

One email per week. New model releases, pricing changes and what they mean for what you build.