Together AI lists Qwen4 Max access at $1.05 per million input tokens
Published API terms give developers a concrete starting point. Alibaba’s claimed task savings remain a separate measure from Together’s token prices.
Loading page…
Published API terms give developers a concrete starting point. Alibaba’s claimed task savings remain a separate measure from Together’s token prices.
Listen to this story
Developers can now evaluate Alibaba’s Qwen4 Max through Together AI’s serverless chat-completions API, with a one-million-token context window and text output. Alibaba claims the model uses 35% fewer tokens and cuts average task costs by about 60% versus Qwen3.8 Max on its Work360 evaluations, but those figures are company-reported benchmark results—not discounts to Together’s token rates. The listing dates the model’s release to November 15, 2026, while leaving the start date for Together API access unspecified.
Together lists rates of $1.05 per million input tokens, $3.15 per million output tokens, and $0.0525 per million cached input tokens.
Qwen4 Max accepts text and images, returns text, and activates 71 billion of its 2.4 trillion parameters per token.
Alibaba’s Work360 benchmark covers 721 specialized fields; Together says benchmark tables await Alibaba’s full technical materials.
Together AI lists serverless access to Alibaba’s Qwen4 Max at $1.05 per million input tokens, $3.15 for output and $0.0525 for cached input. Those published terms give developers a concrete starting point for evaluating the model. Alibaba’s promise of cheaper completed work rests on a separate measure: its own task evaluations.
Together’s listing dates the model’s release to November 15, 2026, without specifying when its own API access began. Developers can use the endpoint Qwen/Qwen4-Max with a Together API key. The page supplies request examples in cURL, Python and TypeScript through the chat-completions API.
Together’s standard rate for input supplied to the model.
Together’s rate for generated output.
The specifications list cached input as a separate pricing category.
Qwen4 Max accepts text and images and returns text. Its one-million-token context window is intended to hold large codebases, document collections and the working history of lengthy agent sessions. Together says deep thinking is enabled by default.
The model uses a sparse mixture-of-experts design: it has 2.4 trillion parameters in total, but activates 71 billion for each token. Alibaba says that active count is 25% below Qwen3.8 Max at the same scale. The model is built on the new Qwen4 architecture.
Alibaba’s efficiency comparison comes from Work360, its in-house benchmark spanning 721 specialized fields. The company says Qwen4 Max outperforms Qwen3.8 Max overall while using 35% fewer tokens in high-effort reasoning mode. It also reports an average cost per task reduction of about 60%.
That figure describes Alibaba’s evaluated tasks, not a 60% reduction in Together’s listed input or output rates. The results are company-supplied; Together’s model page says benchmark tables will arrive with Alibaba’s full technical materials.
For coding work, Together’s model card describes training that pairs software functionality tests with code-security checks. Those checks target weaknesses such as injection and broken access control, aiming to reduce vulnerabilities rather than assess only whether a program completes its intended function.
Loading discussion...
Join the conversation
Explain how you would weigh working code against hidden vulnerabilities.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.