Together Raises Qwen3.7-Max Prices 67%: What to Do

Together AI raised the serverless price of Qwen3.7-Max on November 7, 2026, from $1.50 to $2.50 per million input tokens and from $4.50 to $7.50 per million output tokens, with cached input moving from $0.30 to $0.50. That is a 67% increase on every line, and it puts Together level with Alibaba's own Singapore list price and about 52% above Alibaba's US, Frankfurt and other global regions.

If Qwen3.7-Max carries real agent traffic on Together, your code keeps working but your bill does not. Re-price the workload this week, compare the cheaper routes below, and decide whether to move or to test the newer Qwen3.8 Max instead.

What changed on November 7

Together's pricing page and its serverless models list now show Qwen/Qwen3.7-Max at $2.50 input, $0.50 cached input and $7.50 output per million tokens. The previous rates come from PricingSaaS Pulse, which captured the page on November 6 and November 7 and recorded the jump. The model ID did not change, so there is no redirect, no new endpoint and no code change: the same requests simply cost more from the switch-over onward.

We found no entry about the change in Together's changelog, whose latest entry when we checked was November 5 (the Kimi K3.1 Pro serverless launch). If you budget from an invoice, an internal spreadsheet or an older blog post, the live pricing page is now the only number that matters.

Why the timing matters

Three days before the increase, on November 4, Together removed Qwen/Qwen3.8-2.4T-A95B from serverless inference. Its deprecations page names Qwen/Qwen3.7-Max as the recommended replacement and marks the Qwen3.8 model as not available on on-demand dedicated endpoints. Teams that followed that advice moved onto a model that got 67% more expensive almost immediately.

Together's published lifecycle policy is fairly strict about model swaps. It will not silently redirect a model ID to a release with a price increase above 10%, and it promises email notice and a removal date for serverless deprecations. Those rules cover new models and removals, though, not a price change on a model that stays in the catalog. The practical lesson is the same one we drew from prompt caching economics: treat a third-party per-token rate as a moving input, not a constant.

What Qwen3.7-Max costs elsewhere

Qwen3.7-Max is Alibaba's flagship Qwen3.7 model, and Alibaba's own Model Studio page prices it by region. OpenRouter also lists it, routing to Alibaba endpoints. Prices as of November 8, 2026:

Alibaba lists a 1,000,000-token context window and up to 131,072 output tokens for qwen3.7-max, and its Beijing region also offers batch inference at half price. Together's models table does not list a context length for Qwen3.7-Max, so confirm your longest prompts fit before you rely on either host for long-context agents. Alibaba's full model pricing page also notes that its listed rates exclude limited-time promotions shown in the console.

What the increase does to a real bill

Here is an illustrative month for a coding or support agent: 200 million input tokens and 20 million output tokens, with no caching. These are list-price calculations, not measurements, and they ignore taxes, volume discounts and platform fees.

On that workload the increase adds $260 a month on Together, and moving the same traffic to an Alibaba global region would save about $221 a month against Together's new rate. Caching narrows or widens the gap depending on your hit rate, so plug your own numbers in. If you already route by task, our guide to cost per successful task shows why a cheaper model that fails more often can still cost more.

Should you jump to Qwen3.8 Max instead?

Possibly. Alibaba's model pricing page lists qwen3.8-max and the qwen3.8-max-0902 snapshot at $2 input and $6 output in Singapore, and at $1.65 and $4.951 in its Global and Beijing scopes. In Singapore the newer model is therefore cheaper than Qwen3.7-Max, and elsewhere it costs the same. OpenRouter lists Qwen3.8 Max 0902 at $2 and $6.

The catch is that Together no longer serves the open Qwen3.8 flagship on serverless, so choosing Qwen3.8 Max today means using Alibaba directly or a router. It is also a different model, with different behavior and output length, so treat it as a migration rather than a price swap. Our explainer on open-weight coding models covers how to compare Qwen, Kimi, GLM and DeepSeek without trusting a single leaderboard.

How to make the next price change cheap to absorb

The teams least hurt by this kind of change are the ones that can move a model to a new host with a config change and a test run. Four habits help:

• Keep the base URL and model ID in configuration, not in code. Together, Alibaba Model Studio and OpenRouter all offer OpenAI-compatible endpoints, so the switch is mostly environment variables.
• Put a gateway or router in front. A gateway with fallbacks and spend tracking lets you shift traffic by percentage instead of all at once. See our AI gateways comparison and the OpenRouter tool page.
• Watch the pricing page, not the announcement feed. This change did not show up in Together's changelog. A weekly diff of the providers you depend on, or a budget alert, would have caught it.
• Re-run your evals before cutover. Together's own migration steps recommend an A/B test on representative traffic, then a canary and gradual cutover, because replacement models and hosts rarely behave identically.

A minimal pattern, shown for illustration only (we did not run it against these providers in this session):

# Illustrative: provider chosen by environment, not hard-coded
import os
from openai import OpenAI

client = OpenAI(
base_url=os.environ["LLM_BASE_URL"], # your provider's OpenAI-compatible endpoint
api_key=os.environ["LLM_API_KEY"],
)
resp = client.chat.completions.create(
model=os.environ["LLM_MODEL"], # e.g. Qwen/Qwen3.7-Max on Together, qwen3.7-max on Model Studio
messages=[{"role": "user", "content": "Summarize this diff"}],
)

Note that model IDs differ by host (Qwen/Qwen3.7-Max on Together, qwen3.7-max on Model Studio, qwen/qwen3.7-max on OpenRouter), so keep a small mapping table per provider.

Staying on Together: pros and cons

Reasons to stay

• No migration work: same model ID, same API, same SDK.
• One account and one bill for many open models, including Kimi K3.1 Pro ($2 input, $10 output), GLM-5.3 ($1.40, $4.40) and DeepSeek V4.1 Flash ($0.30, $1.20), per Together's pricing page on Nov 8, 2026.
• A path to dedicated endpoints, fine-tuning and post-training on the same platform, plus new Anthropic Messages and OpenAI Responses API support for some models (added Nov 3).

Reasons to move

• For Qwen3.7-Max specifically, Together is now no cheaper than Alibaba's most expensive region and about 52% more than its global regions.
• The change landed without a changelog note, three days after Together pointed Qwen3.8 users to this model.
• If you need the newer Qwen3.8 flagship, Together no longer serves it on serverless.

Verdict

If Qwen3.7-Max is a meaningful line on your bill, re-run your cost model now. For US or EU traffic, Alibaba Model Studio's global regions list the same model about 34% below Together's new rate, and OpenRouter's default Alibaba route is lower still before its credit fee. Test quality and latency on your own prompts before moving, because hosts are not interchangeable.

If you value one bill across many open models, staying on Together AI is defensible, but put a gateway with fallbacks in front and add Qwen3.8 Max to your next eval run. Either way, check the Qwen3.7-Max price on the live page before every budget review.

Sources

• Together AI: Pricing (accessed Nov 8, 2026)
• Together AI docs: Serverless models
• Together AI docs: Deprecations and model lifecycle policy
• Together AI docs: Changelog
• Alibaba Cloud Model Studio: qwen3.7-max model info
• Alibaba Cloud Model Studio: Model pricing
• OpenRouter: Qwen3.7 Max
• PricingSaaS Pulse: Together raised Qwen3.7-Max inference prices (Nov 7, 2026)