Together reverses course, cuts Qwen inference prices 40% two days after a hike
By Rob Litterst, Co-founder, PricingSaaS
Published
Together’s pricing page · Change detected
What changed
Together cut serverless inference pricing on two Qwen models: Qwen3.8 Flash fell from $0.15/$0.47 to $0.09/$0.28 per 1M input/output tokens, and Qwen3.7-Max fell from $2.50/$7.50 to $1.50/$4.50 — both roughly 40% reductions. Cached-token pricing on Qwen3.7-Max moved the other way, rising from $0.25 to $0.30 per 1M tokens.
Why it matters
Inference hosting for popular open-weight models like Qwen, DeepSeek, and Kimi is looking increasingly commodity-priced, with providers competing on thin, fast-moving margins rather than sticky list prices. Watch whether this cut holds or gets reversed again within days, and whether the quiet rise in cached-token pricing — up even as headline prices fell — becomes where Together rebuilds margin.
In context
This cut lands just two days after Together raised Qwen3.7-Max pricing 25% to $2.50/$7.50 on Sep 21, the tail end of a stretch that saw the same model hiked repeatedly through early September before today's reversal. That whiplash fits Together's overall cadence — 36 changes in the past year, roughly one every 7 days — making it the fastest-moving page in the credits-and-usage-billing pattern that touched 348 companies industry-wide over six months. Unlike peers such as Snyk restructuring Enterprise onto credit-based pricing or Notion ending its free Workers beta ahead of October credit billing, Together isn't repackaging — it's adjusting raw per-token unit prices near-weekly, treating inference cost like a spot market tied to GPU economics and rival hosts of the same open-weight models.