Together AI has raised the price of running Alibaba’s flagship Qwen3.7-Max by two-thirds, without announcing the change. Serverless input tokens now cost $2.50 per million, up from $1.50, with cached input at $0.50 and output at $7.50, also up 67 percent. The new prices match Alibaba’s own Singapore list price, and they sit about 52 percent above what Alibaba charges in the United States and other global regions. The hike, which took effect on November 7, was spotted by a pricing tracker, not announced in Together’s changelog.
What changed
Every line of the price rose by the same share. Input tokens went from $1.50 to $2.50 per million, cached input from $0.30 to $0.50, and output tokens from $4.50 to $7.50. Together’s own pricing page, checked on November 8, lists the new figures for Qwen3.7-Max on serverless inference.
A capture of the same page from November 6 shows the old ones, so the change is real; only its announcement is missing. There is no entry about the rise in Together’s changelog, whose latest entry when checked was November 5, for the Kimi K3.1 Pro serverless launch.
Silent repricing is not unusual for inference providers, whose costs move with GPU demand and capacity pressure. But a 67 percent jump on a flagship model, with no note to customers at all, is a sharp way to do it.
The context
The hike follows a quieter change three days earlier. On November 4, Together removed Qwen3.8-2.4T-A95B, the previous flagship, from serverless inference, and its deprecations page names Qwen3.7-Max as the recommended replacement. Developers steered from the old flagship to the new one now find the replacement costs two-thirds more than the price they were paying a week ago.
Alibaba’s own prices make the comparison awkward. Its Model Studio lists Qwen3.7-Max at $2.50 input and $7.50 output in Singapore — exactly Together’s new price — but at $1.65 and $4.951 in the United States, Frankfurt and other global regions.
The older Qwen3.8 Max, meanwhile, lists at $2 input and $6 output in Singapore and the same $1.65 and $4.951 globally, so the previous generation is now cheaper than its replacement on Together. The gap is a reminder that the same model can cost very different amounts depending on where it is served from.
Together’s own menu underlines the premium. Its Qwen3.7-Plus lists at $0.32 input and $1.28 output, and Qwen3.8 Flash at $0.15 and $0.47 — a fraction of the flagship’s new rates. Developers who do not need the largest model were already paying far less, and the hike widens that gap.
OpenRouter offers another way around the hike. It lists Qwen3.7-Max as well, routing requests to Alibaba endpoints, which puts its price closer to Alibaba’s own global rates than to Together’s new ones. For teams already set up with a router, switching providers can be a configuration change rather than a migration.
Our Read
This looks like demand pricing, not a mistake. Qwen3.7-Max is Alibaba’s flagship, and Together’s serverless tier is where smaller teams and hobbyists meet it first. Raising the price to match Alibaba’s Singapore list removes Together’s discount without removing the model, and the missing changelog entry suggests the company would rather have the price discovered than announced.
Developers have options, and that is the real check on the hike. Alibaba’s own US pricing undercuts Together by about a third, and OpenRouter routes the same model through Alibaba endpoints. For teams already paying per token, the arithmetic is simple: a 67 percent rise on a flagship model is worth an afternoon of re-routing, and the providers know it.
Why It Matters
Inference pricing compounds. A model that costs $2.50 per million input tokens instead of $1.50 changes the economics of every app built on it, from chatbots to coding agents. The output price matters more: at $7.50 per million, long generations get expensive fast.
Silent hikes also break trust, because teams budget against published prices and discover the change on the invoice. The wider lesson is that the cheapest way to run a model keeps shifting, and the gap between hosted prices and local costs is part of why running models at home keeps gaining ground.
Background
Qwen3.7-Max is Alibaba’s flagship model in the Qwen3.7 family, with a one-million-token context window and up to 131,072 output tokens. Alibaba lists batch inference at half price in its Beijing region, a discount that looks more attractive now that serverless rates have risen.
Pricing has been a live issue for Qwen all month. QwenWork sold AI tokens in a Singles Day promotion earlier in November, and Alibaba’s regional price gaps have been a standing feature of its Model Studio listings. The Together hike is the latest entry in that running story.
Source: Together AI pricing · price-change tracker
0 Comments