68 / 100

Fireworks AI pricing teardown

https://fireworks.ai/pricing

Fireworks runs a transparent, usage-based pricing page organized by product line (Serverless Inference, Training, On Demand Deployments) rather than good-better-best plans, with public per-token, per-training-token and per-GPU-second rates. It's strong on price visibility and developer self-serve, but the page is dense, offers no free trial or free credits, has no recommended path or FAQ, and the serverless inference rates are mostly deferred to a separate docs page.

Tier structure

Serverless Inference

pay per token (Embeddings from $0.008/1M input tokens)

Training

from $0.50/1M training tokens (LoRA SFT, models up to 16B)

On Demand Deployments

from $8.00/hr per H100/H200 GPU ($0.134/min)

Value metric

usage-based: per token, per training token, per GPU second

How limits scale

Training price per 1M tokens (LoRA SFT, by model size) $0.50 (≤16B) → $3.00 (16-80B) → $6.00 (80-300B) → $10.00 (>300B)
On-demand GPU price per hour $8.00 (H100) → $8.00 (H200) → $13.00 (B200) → $15.00 (B300) → $20.00 (GB300)
Embeddings price per 1M input tokens $0.008 (up to 150M params) → $0.016 (150M-350M) → $0.10 (Qwen3 8B)

Escalation logic

Rather than escalating plans, the three products are separate workloads that map to a maturity path: start serverless with zero setup, customize via training, then move to dedicated GPUs for speed, rate limits and cost at scale. Upgrade triggers are implied by copy (faster speeds, higher rate limits, lower costs at scale) but no tier-to-tier comparison is shown.

Psychological anchors

  • Pay-as-you-go entry anchor — Serverless is positioned first with 'zero setup and no cold starts' and fractions-of-a-cent embedding prices ($0.008/1M tokens), making the entry cost feel negligible.
  • Promotional/free anchor — A 'Limited-time offer' badge and 'Muse Glimmer 30B SFT is free via UI and Serverless Training API until 11/30' with 'No credit card required' create urgency and a free way to try training.
  • Same-price-as-base-model value claim — 'Serve fine-tuned models for the same price as base models' removes a common objection to customizing models.
  • Tiered price escalation by model size — Training prices step up ~2-6x across parameter-count bands, transparently tying cost to compute consumed.
  • Enterprise upsell via 'Contact us' — Dual hero CTAs (Get started / Contact us) and copy like 'faster speeds, lower costs, and higher rate limits' for enterprise deployments, plus a 1.5x premium for region-restricted deployments, anchor a higher-value tier.
  • Spend controls as risk reducer — Pre-paid credits, Auto Reload and a monthly spend limit reassure usage-based buyers that bills won't run away.
  • Trust signals near the footer — SOC 2 Type 2, ISO 27001, 27701 and 42001 badges support enterprise credibility.

What this page is optimizing for

Primarily developer self-serve adoption with an enterprise lead-gen path: the hero says 'Start building in seconds, self-serve' with a primary 'Get started' CTA, while 'Contact us' and the region-restricted/enterprise copy capture larger deployments. Public per-unit prices and pre-paid credits push low-friction entry and expansion through usage.

Red flags

  • Core serverless model pricing (the most popular entry point) isn't shown inline; users are sent to a separate pricing docs page, adding friction right where most visitors start.
  • No recommended or 'most popular' path and no side-by-side comparison, so visitors must self-diagnose which of the three products fits them.
  • Page is very dense: multiple tables, long training-token estimation notes and a growing model catalog make scanning hard and bury key numbers.
  • No free trial or free starter credits for core inference; entry requires adding a payment method to buy credits (the free training offer is time-limited and model-specific).
  • Pure usage-based pricing makes monthly cost hard to predict, and there is no calculator or example-cost scenarios to help estimate it.
  • No annual or committed-use discount is visible, and enterprise pricing is hidden behind 'Contact us' with no indication of volume tiers.
  • Several serverless training model names and prices are hard to compare (differing context, prefill, cached, sample, train columns) without a plain-language explanation.

Best-practices scorecard

What works · 3

  • Visible prices — Training, embeddings and GPU rates are all published; only the serverless model rates are pushed to a docs link.
  • Value metric matches usage — Per token, per training token and per GPU second map directly to how customers consume the product, with 'no extra charges for start-up times'.
  • Trust signals present — SOC 2 Type 2 and ISO 27001/27701/42001 certifications appear on the page.

Half measures · 3

  • Tier differences are scannable — The three-card overview is clean, but the detailed sections are table-heavy and lack a cross-product comparison.
  • FAQ or objection handling — Objections are handled inline (same price for fine-tuned models, spend limits, token estimation notes) but there is no dedicated FAQ section.
  • Clear CTAs per tier — Each card has a 'See pricing' link and the hero has Get started/Contact us, but the product cards lack direct signup CTAs.

What's missing · 2

  • Clear recommended tier — None of the three products is highlighted or badged as the default starting point beyond the implicit order.
  • Annual discount offered — No annual or committed-use discount is shown; billing is pre-paid usage credits.