Price Per Token lists Qwen3.5-397B-A17B with lower per-token API prices than Alibaba Cloud’s Model Studio page, while Qwen’s Hugging Face model card describes the model as a 397B-parameter, 17B-activated Apache-2.0 release and links it to the hosted Qwen3.5-Plus variant.
Price Per Token updated its Qwen3.5-397B-A17B listing on June 19, 2026, adding API price, context-window, and benchmark details for Alibaba’s large mixture-of-experts model.
Price Per Token lists Qwen3.5-397B-A17B at $0.385 per 1 million input tokens and $0.900 per 1 million output tokens. The same listing says the model has a 262K-token context window, is available through the Qwen API, and includes benchmark metadata such as a GPQA score of 86.1 and an intelligence-score percentile.
The official Qwen model card on Hugging Face identifies Qwen3.5-397B-A17B as a model with 397 billion total parameters and 17 billion activated parameters. Qwen’s Hugging Face page also states that the model is released under the Apache-2.0 license and says Qwen3.5-Plus is the hosted version corresponding to Qwen3.5-397B-A17B.
Alibaba Cloud’s Model Studio pricing page gives a different price for international access. For qwen3.5-397b-a17b requests in the 0 to 256K token range, Alibaba Cloud lists $0.60 per 1 million input tokens and $3.60 per 1 million output tokens.
The difference between the Price Per Token listing and Alibaba Cloud’s Model Studio page is material for developers comparing production costs. On input tokens, Alibaba Cloud’s listed international Model Studio price is higher than the Price Per Token figure. On output tokens, the gap is larger: Alibaba Cloud lists $3.60 per 1 million output tokens, compared with $0.900 per 1 million output tokens on Price Per Token.
The sources do not, by themselves, explain the discrepancy. Possible differences could include the access channel, region, account type, promotional terms, routing, or update timing, but those explanations are not stated in the provided source excerpts. For procurement or production budgeting, Alibaba Cloud’s own pricing page should be treated as the primary reference for Model Studio charges, while the Price Per Token entry is useful as a third-party comparison point that should be checked against provider documentation before use.
The Qwen Hugging Face model card’s parameter description places Qwen3.5-397B-A17B among very large mixture-of-experts systems: 397 billion total parameters with 17 billion activated per inference step. That architecture can allow a model to have a large overall parameter pool while activating a smaller subset for each token, though the provided sources do not give enough detail to compare its latency, throughput, or serving economics against other models.
The Apache-2.0 license noted by Qwen on Hugging Face is also significant for developers evaluating self-hosting or integration options. However, API pricing and hosted availability are separate from the open model card: Qwen’s model card connects the model to Qwen3.5-Plus as the hosted version, while Alibaba Cloud’s Model Studio page provides the cited international per-token charges.
The current source set supports three cautious conclusions: Qwen identifies Qwen3.5-397B-A17B as a 397B-total, 17B-activated Apache-2.0 model; Price Per Token reports lower API rates and benchmark metadata for the model as of June 19, 2026; and Alibaba Cloud’s own Model Studio page lists higher international prices for qwen3.5-397b-a17b in the 0 to 256K token band.
Until the difference between the third-party listing and Alibaba Cloud’s pricing page is clarified, developers should verify the final rate in the provider console or contract terms before estimating deployment costs.
Price Per Token updated its Qwen3.5 397B A17B listing on June 19, 2026, adding API price, context window, and benchmark details for Alibaba’s large mixture of experts model.
What the listings say Price Per Token lists Qwen3.5 397B A17B at $0.385 per 1 million input tokens and $0.900 per 1 million output tokens.
The same listing says the model has a 262K token context window, is available through the Qwen API, and includes benchmark metadata such as a GPQA score of 86.1 and an intelligence score percentile.
Continue reading
