Fireworks

Fireworks added GLM 5.3 Flash as a new base model in the Serverless Training API pricing table, with its own 200K context window and per-1M-token pricing for prefill, cached prefill, sample, and train.

Serverless Training API model catalog updated: Qwen 3.5 9B (64k, $0.66 prefill) removed and Qwen 3.6 27B replaced by Qwen 3.8 27B at identical pricing; DeepSeek V4 Flash 0731 (262K) and Muse Glimmer 30B (128K) added with per-token prefill/cached/sample/train pricing.

On-Demand GPU pricing rising from Sep 1: H100/H200 $7.00→$8.00, B200 $10.00→$13.00, B300 $12.00→$15.00, GB300 $18.00→$20.00/hr; region surcharge increased 1.25x→1.5x.

Serverless Training API: Qwen 3.6 27B context window increased from 65k to 128k tokens (prefill/sample/train prices unchanged). On-Demand Pricing added a new GB300 288GB GPU tier at $18.00/hour and a new note that region-restricted deployments (US, Europe) are priced at 1.25x the listed rates.

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.

On-Demand GPU pricing increased: H100 $6.00→$7.00/hr, H200 $6.00→$7.00/hr, B200 $9.00→$10.00/hr, new B300 tier at $12.00/hr. Also, the detailed per-model serverless pricing table was replaced with a docs-page link.

Serverless pricing tier renamed from 'Turbo' to 'Fast'. The 'Turbo + Priority tiers' description on the pricing page now reads 'Fast + Priority tiers'.

Serverless Text and Vision model prices removed from pricing page and redirected to docs.fireworks.ai/serverless/pricing. All per-model token prices (DeepSeek, Kimi, GLM, OpenAI gpt-oss, MiniMax) no longer publicly listed. On-Demand GPU prices increased as previously announced (May 1, 2026): H100 $6.00→$7.00/hr, H200 $6.00→$7.00/hr, B200 $9.00→$10.00/hr, B300 $11.00→$12.00/hr. Advance notice note removed.

Speech to Text and Image Generation sections removed from pricing page; DeepSeek-V4-Pro added at $1.74/$0.145 cached/$3.48 per 1M tokens.

New On-Demand GPU tier B300 288 GB added to on-demand deployments pricing at $11.00/hour billed per second. Fine tuning pricing expanded from 2 columns (Supervised Fine Tuning, DPO) to 4 columns (LoRA SFT, LoRA DPO, Full Param SFT, Full Param DPO) with new Full Param prices added per model tier.

Streaming ASR v1/v2 removed from STT pricing. DeepSeek V3, Kimi K2 Instruct/Thinking, and gpt-oss-20b lost cached pricing. Caching footnote updated to 50% universal rate. H100 80 GB GPU on-demand price increased from $4.00/hr to $6.00/hr (50% increase). MiniMax M2 family cached token price decreased from $0.15 to $0.03 per 1M tokens.

Added Kimi K2.5 Turbo model with pricing at $0.99 input, $0.16 cached input, $4.94 output per 1M tokens

Added policy note explaining fine-tuning with reasoning traces increases tuned tokens due to conversation unrolling, with reference to SFT documentation

Qwen3 235B Family removed from serverless inference pricing [was $0.22/1M input, $0.88/1M output tokens] GLM-5 added to serverless inference pricing at $1.00/1M input, $0.20 cached input, $3.20/1M output tokens

Cached pricing display restructured: per-row cached prices removed from all model rows. Unified footnote now states cached input tokens are 50% for all text and vision models. Previous Qwen3 VL series exception removed.

Kimi K2 Thinking model added alongside existing Kimi K2 Instruct, both with same pricing $0.60 input, $0.30 cached, $2.50 output Qwen3 235B Family and GLM-4.5 Air renamed to Qwen3 235B Family (GLM-4.5 Air component removed)

FLUX.1 Kontext Pro image model added at $0.04 per image H200 141 GB GPU price decreased from $6.99/hour to $6.00/hour (14% decrease)

Developer plan tier removed from main pricing section (replaced with product-based presentation) Qwen3 30B model added at $0.15 input / $0.60 output per 1M tokens

DeepSeek R1 (Basic) model added with pricing $0.55 input, $2.19 output per 1M tokens Meta Llama 4 Scout (Basic) model added with pricing $0.15 input, $0.60 output per 1M tokens