Fireworks: New Serverless Training API added (Qwen 3.5 9B, 3.6 27B, Kimi K3)
What happened
Between the Jul 27, 2026 and Aug 3, 2026 captures of Fireworks's pricing page, Pulse recorded one change: Serverless Training API introduced — per-token LoRA training pricing for Qwen 3.5 9B ($0.66 prefill/$1.463 train), Qwen 3.6 27B ($1.86/$4.103), and Kimi K3 ($10.87/$32.55) per 1M tokens. In detail: Added Serverless Training API: LoRA training on a shared always-on trainer pool for Qwen 3.5 9B, Qwen 3.6 27B, and Kimi K3, priced per 1M tokens for prefill/cached-prefill/sample/train. Pulse files it as a packaging change (new add-on) with an upward direction. Pulse's latest capture (Oct 19, 2026) lists On-Demand at $13 per GPU hour (B200 180GB) + $15 per GPU hour (B300 288GB), Serverless at $0.08 per 10M input tokens (embeddings, up to 150M params) + $0.16 per 10M input tokens (embeddings, 150M-350M params) and Fine Tuning at $0.5 per 1M training tokens, LoRA SFT (models up to 16B params).
It lands 8 days after Fireworks's previous change, "Fireworks: new Serverless Training API for pay-per-token LoRA training" (Jul 28, 2026). Before that Pulse recorded "Fireworks: On-Demand GPU pricing increased, new B300 tier added" (Jul 13, 2026), "Fireworks renames Serverless tier from Turbo to Fast" (Jun 8, 2026) and "Fireworks hikes GPU prices, moves model pricing behind docs wall" (May 12, 2026). Fireworks has been tracked since 2024-11-17.
In the same window Sketch, Adobe and Corel also changed their Marketing & Content pricing pages. Those moves are listed further down, and the Pulse Agent can compare them side by side.
- Serverless · Add-onAdded
BeforeNo Serverless Training API
AfterServerless Training API
| Plan | Field | Before | After | Change |
|---|---|---|---|---|
| Serverless | Add-on | No Serverless Training API | Serverless Training API | Added |
Values as written on Fireworks's public pricing page in the two captures Pulse compared; where the page states no number, the cells give the wording before and after.
Evidence
The two captures Pulse comparedFull-page captures of Fireworks's pricing page (fireworks.ai) from Jul 27, 2026 and Aug 3, 2026; the change is marked around the middle of the page, where Serverless Training API now appears.
Sign in to see the evidence
The two full-page captures with this change marked up are free with an account.
Fireworks, the last 90 days
Full history- Sep 1, 2026improvedFireworks adds DeepSeek V4 Flash, Muse Glimmer 30B to training API
- Aug 17, 2026increaseFireworks: On-Demand GPU pricing up 11-30% from Sep 1; region surcharge to 1.5x
- Aug 11, 2026improvedFireworks: New GB300 GPU tier ($18/hr) + region premium pricing
- Aug 5, 2026packagingFireworks: New Serverless Training API added (Qwen 3.5 9B, 3.6 27B, Kimi K3) · this change
- Jul 28, 2026packagingFireworks: new Serverless Training API for pay-per-token LoRA training
- Jul 13, 2026increaseFireworks: On-Demand GPU pricing increased, new B300 tier added
- Jun 8, 2026packagingFireworks renames Serverless tier from Turbo to Fast
- May 12, 2026increaseFireworks hikes GPU prices, moves model pricing behind docs wall
Across Marketing & Content in the same window
All Marketing & Content changesQuestions about this change
Did Fireworks raise prices in August 2026?
No list-price increase was recorded between Jul 27, 2026 and Aug 3, 2026. The change Pulse detected was a packaging change: new add-on.
What changed on Fireworks's pricing page between Jul 27, 2026 and Aug 3, 2026?
Serverless add-on: No Serverless Training API → Serverless Training API, added. Pulse compared the captures it took on Jul 27, 2026 and Aug 3, 2026; the before/after table above lists each item as written on the page.
How often does Fireworks change its pricing?
Pulse recorded 4 other pricing-page changes for Fireworks in the 90 days before this one; the previous change was "Fireworks: new Serverless Training API for pay-per-token LoRA training" on Jul 28, 2026; Fireworks has been tracked since 2024-11-17.