Fireworks AI Pricing in 2026: Every Serverless Model Priced

Fireworks AI prices 12 models on its self-serve serverless rate card as of 2026-11-03, spanning $0.05/$0.20 per million tokens for NVIDIA Nemotron 3.5 Lightning 30B A3B to $3.00/$15.00 for Kimi K3 — and its $0.006 cached-input rate on DeepSeek V4.1 Flash is the deepest published cache discount among the hosts DeAI tracks.

Key facts

• The serverless card is 12 rows, priced input/cached-input/output: Kimi K3 at $3.00 / $0.30 / $15.00, Qwen 3.8 Max at $2.00 / $0.25 / $6.00, GLM 5.3 at $1.40 / $0.26 / $4.40, Muse Spark Open at $1.25 / $0.15 / $4.25, DeepSeek V4.1 Flash at $0.30 / $0.006 / $1.20, MiniMax M3 at $0.30 / $0.06 / $1.20, gpt-oss-120b at $0.15 / $0.015 / $0.60, GLM 5.3 Flash at $0.15 / $0.03 / $0.50, and Nemotron 3.5 Lightning 30B A3B at $0.05 / $0.01 / $0.20 (Fireworks pricing, accessed 2026-11-03).
• Every open-weight row carries a cache discount: GLM 5.3 Flash caches at 80% off, gpt-oss-120b and MiniMax M3 at 90% off, and DeepSeek V4.1 Flash at 98% off ($0.006 versus $0.30) — cached input "bills at half price or less, with no write fee" per the pricing page (Fireworks pricing).
• Mode levers stack: Batch mode bills at 50% of Standard on input and output, Spot Serverless takes 30% off Standard uncached prices for retry-tolerant traffic, and Priority runs at a 25% premium — all on the same endpoint (Fireworks Serverless Modes).
• Rate limits are adaptive and per model: small models (under 800B parameters) reach ceilings of 216M Total Prompt TPM, 54M Uncached TPM, 2.16M Generated TPM at the top spending tier, with +25% steps every 15 minutes and a 72-hour idle reset (Fireworks rate limits).
• Third-party speed: Artificial Analysis measures DeepSeek V4.1 Flash (non-reasoning) at 418 tokens/s — the fastest of Fireworks's 12 listed models — and the provider's cheapest blended price is $0.04/M on Nemotron 3.5 Lightning BF16 (Artificial Analysis, accessed 2026-11-03).
• Retention is policy, not attestation: the privacy notice (last modified September 16, 2026) states no training on prompts without explicit opt-in and no logging of prompt or generation data for open models without opt-in; the Response API retains conversation data up to 30 days when store=True (Fireworks privacy notice; Zero Data Retention docs).

What Fireworks AI is

Fireworks AI is a Redwood City inference provider that serves open-weight and closed models per-token on shared serverless capacity, by GPU-second on On Demand deployments, or via Reserved Throughput — dollars-per-minute reservations backed by a 99.9% uptime SLA. The same endpoint speaks Standard, Priority, Fast, Batch, and Spot modes, which is the structural difference from hosts like Cerebras, whose card is fixed: on Fireworks, the mode you select per request changes the bill. The platform also sells GPU time (H100/H200 at $8.00/hour, B200 at $13.00, B300 at $15.00, GB300 at $20.00) and trains models, but this page prices the serverless token card, which is what most builders touch.

The company's footing: Fireworks raised a $52 million Series A in 2023 led by Sequoia, then a $100 million Series B in July 2024 at a $552 million valuation co-led by Sequoia and Lightspeed (Reuters). Recent quarters have added Microsoft Foundry as a first-party sales channel — usage counts toward Azure Consumption Commitments — and a Sovereign offering that runs Fireworks-managed inference inside a customer's own cloud account or data center, in private enterprise preview.

The full rate card, priced (2026-11-03)

Every open-weight serverless model, Standard mode, first-party list rates accessed 2026-11-03. Closed models included for comparison (Ember-1, Rune):

Two structural notes the table hides. First, US-only capacity is a 50% premium on every model except Muse Spark Open, where it's waived — DeepSeek V4.1 Flash (US) runs $0.45/$0.009/$1.80 versus $0.30/$0.006/$1.20. That's the price of a US-residency guarantee, and it's the cheapest one on this card by a wide margin. Second, models not on the card are priced by size: $0.10 per million tokens under 4B parameters, $0.20 for 4B–16B, $0.90 above 16B, and $0.50–$1.20 for MoE architectures — with one flat rate covering input and output and no separate cached-input discount. A fine-tune you upload therefore lands on a predictable, if undifferentiated, price.

Priority mode adds 25% across the board ($3.75/$0.375/$18.75 on Kimi K3, $0.375/$0.0075/$1.50 on DeepSeek V4.1 Flash) in exchange for peak-load priority; Kimi K3 Fast runs a further 50% above Standard at $4.50/$0.45/$22.50. A decision-oriented row exists too: Rune 26B-A4B, a "System One" model that never generates output tokens, bills $0.042 input / $0.0084 cached / zero output — you pay for input only.

The cache economics are the story

Fireworks's cached-input pricing is the most aggressive published schedule among the hosts DeAI has priced this quarter. The FAQ states caching is automatic — no setup, no write fee — with cached input at "50% off by default, with deeper discounts on select models." The rate card shows how deep: 80% off on GLM 5.3 Flash, 90% off on gpt-oss-120b and MiniMax M3, and 98% off on DeepSeek V4.1 Flash, whose $0.006 cached rate undercuts every DeepSeek host price DeAI has recorded to date.

Compare the field on the same checkpoints: Groq bills cached input at 50% off with a 2-hour expiry on its GPT-OSS trio; SambaNova's MiniMax rows cache at 90% off; Cerebras bills cached tokens at the standard input rate with no discount at all. For agentic workloads — long system prompts, tool definitions, and multi-turn context that repeat every request — a 90–98% cache discount converts directly into a fraction of the bill. Fireworks adds session affinity, which keeps each agent pinned to the server already holding its context; the company says long runs reach cache hit rates up to 99.6% — a vendor claim, not a number we have independently verified.

Speed: the other side of the trade

Third-party medians put Fireworks in the mid-pack on speed. On Artificial Analysis's provider page, DeepSeek V4.1 Flash (non-reasoning) tops the provider's own catalog at 418 tokens/s with a 0.53s time to first token, Nemotron 3.5 Lightning (BF16) follows at 399 tokens/s, and MiniMax M3 at 290 tokens/s; Kimi K3 (low) runs 114 tokens/s (Artificial Analysis, accessed 2026-11-03). On gpt-oss-120b specifically — the model DeAI tracks across 16 providers — Fireworks's endpoint does not appear in the published top tier, which Cerebras (1,684 tokens/s), SambaNova (709), and Groq (473) dominate. Fireworks is selling breadth, caching, and mode flexibility; if raw tokens-per-second is the requirement, the wafer-scale and LPU hosts remain the specialists.

One pricing subtlety from the same source: Fireworks's gpt-oss-120b listing on Artificial Analysis shows a blended price around $0.20 per million tokens — roughly 4x CoreWeave's $0.04 floor for the same checkpoint — so the cache discount, not the base rate, is where Fireworks's cost advantage lives for cache-heavy traffic.

Privacy: a policy, not an attestation

Fireworks's privacy notice (last modified September 16, 2026) states three commitments up front: no AI training on your data without explicit opt-in, user control over access/download/deletion, and zero data retention — "We do not log or store prompt or generation data for any open models without explicit user opt-in." The Zero Data Retention documentation is more specific: prompt and generation data exist only in volatile memory for the duration of the request (including cache-resident prompt data for several minutes when caching is active), metadata such as token counts is logged as needed to deliver the service, and users can opt in to logging for advanced features like FireOptimizer.

Two exceptions matter. The Response API retains conversation data — prompts, responses, tool calls — for up to 30 days when store=True, which is the default; you disable it with store=False or delete immediately via the response ID. And enterprise account admins can enable an additional ZDR policy that rejects any request or job that would persist customer content. DeAI treats these as self-published policy statements, not independently verified guarantees — the distinction that separates a policy from an attestation — but as published terms go, the open-model default (nothing persisted without opt-in) is among the clearest on the tracked field. The pricing page also lists SOC 2 Type II, ISO 27001, 27701, and 42001, and HIPAA with a BAA.

Rate limits, exactly

Fireworks's limiter is adaptive, per account and per model, in tokens per minute — three buckets (Total Prompt TPM, Uncached Prompt TPM, Generated TPM), each growing independently. Base ceilings depend on the model's total parameter count and your spending tier; the published maximums at Tier 3:

The adaptive mechanics: when your recent one-minute usage exceeds 50% of your current limit, the limit rises 25%, at most once every 15 minutes, toward the base ceiling — a full ramp takes a few hours. Limits shrink 20% per step after your highest sampled 15-minute usage stays below 33% of the limit for 3 days (never below the lower bound), and reset to the starting default after 72 hours of no traffic. Eligible accounts get Elastic Burst, temporary capacity above the ceiling billed at normal serverless rates with no burst premium. Hitting a limit on one model does not affect your traffic to any other model. Reserved Throughput customers get guaranteed capacity backed by the SLA, and overage above a reservation bills at standard Serverless rates — never a penalty rate.

The cost math, worked

For an agentic workload with a 20k-token system prompt plus tools repeated per request and a 2k-token response: at a 95% cache-hit rate on DeepSeek V4.1 Flash, each request bills roughly 1,000 uncached input tokens ($0.0003), 19,000 cached tokens ($0.000114), and 2,000 output tokens ($0.0024) — about $0.0028 per request in Standard mode, or $0.0014 in Batch. The same request uncached would bill $0.0030 + $0.0024 = $0.0054, so the cache schedule is worth roughly half the bill at this shape. Your workload's prompt-reuse profile decides how much of this transfers; measure your own hit rate from the Serverless dashboard's historical limit samples and usage exports before comparing hosts on list price alone.

How to choose

Fireworks is the fit when you need many open models behind one OpenAI-compatible endpoint, cache-heavy agentic traffic, or mode-level cost control (Batch for anything that can wait, Spot for retry-tolerant jobs) without a commitment. It is not the fit when a single model must run at maximum tokens-per-second — that's Cerebras or SambaNova territory — or when the absolute floor price per million tokens on commodity GPU capacity is the only criterion — DeepInfra and CoreWeave hold that end. For US-residency-constrained workloads, Fireworks's 50% US-only premium is the cheapest published residency uplift among the hosts priced in this series, but verify the residency contract terms against your own compliance requirements before relying on it.

Current state (November 2026)

The card reflects a fast-moving catalog: DeepSeek V4.1 Flash 1017 is listed as a separate row from the base Flash (same price), Nemotron 3 Ultra is still labeled Preview, Qwen 3.8 Max and Ember-1 sit at the premium end, and Spot Serverless is available only on select models (GLM-5.3, Kimi K3, and both DeepSeek V4.1 Flash variants on the published spot table). Prices move — the price-index deltas DeAI has published through October 2026 record repeated repricings across this field — so treat every figure here as dated to 2026-11-03 and re-verify against Fireworks pricing before committing spend.

FAQ

How much does Fireworks AI cost per million tokens?

As of 2026-11-03, the serverless rate card runs from Nemotron 3.5 Lightning 30B A3B at $0.05 input / $0.20 output to Kimi K3 and Ember-1 at $3.00 / $15.00, with GLM 5.3 at $1.40 / $4.40, DeepSeek V4.1 Flash at $0.30 / $1.20, and gpt-oss-120b at $0.15 / $0.60 in between. Models not individually listed are priced by parameter count ($0.10 under 4B to $0.90 above 16B, flat across input and output).

What is the cheapest way to run DeepSeek V4.1 Flash on Fireworks?

Standard mode lists $0.30 / $1.20 with cached input at $0.006 (98% off). Batch mode bills at 50% of Standard ($0.15 / $0.60), and Spot Serverless takes 30% off uncached Standard ($0.21 / $0.84) — stack both for a floor around $0.105 / $0.42 per million tokens for retry-tolerant background jobs.

What are Fireworks AI's rate limits?

Per-model adaptive limits in TPM: small models (under 800B) reach 216M Total Prompt TPM / 54M Uncached / 2.16M Generated at the top spending tier; 800B–1.6T models 43.2M / 10.8M / 432k; models at 1.6T and above 21.6M / 5.4M / 216k. Limits grow 25% at most once per 15 minutes when usage runs above 50% of the current limit, shrink 20% per step after 3 days below 33% utilization, and reset after 72 idle hours. Elastic Burst can lift eligible accounts above the ceiling at no premium.

Does Fireworks AI offer prompt caching?

Yes — automatic, no setup, no write fee, cached input at 50% off by default with deeper discounts on select models (80% off GLM 5.3 Flash, 90% off gpt-oss-120b and MiniMax M3, 98% off DeepSeek V4.1 Flash on the published card). Session affinity pins each agent to the server holding its context; Fireworks claims long-run hit rates up to 99.6%.

Does Fireworks AI train on your prompts?

Per its privacy notice (September 16, 2026), no — not without explicit opt-in — and it states it does not log or store prompt or generation data for any open models without opt-in, with data held only in volatile memory per request. The Response API retains conversation data up to 30 days when store=True (the default) unless you set store=False. These are self-published policy statements, not independent attestations.