Qwen3.5-Plus: Two Snapshots, a 1M Window and a 256K Price Step

Qwen3.5-Plus looks like one model and is actually three decisions. Which snapshot to pin, whether your prompts cross a pricing threshold, and whether you want the hosted product at all when the open weights underneath are Apache 2.0. Getting those right matters more than any benchmark table.

The model itself is well defined. Qwen's model card calls Qwen3.5-Plus "the hosted version corresponding to Qwen3.5-397B-A17B", served on Alibaba Cloud Model Studio with a 1M default context and built-in tools. It takes text, images and video, and returns text.

Plus is the hosted 397B, with extras

Underneath is a 397B mixture-of-experts model with about 17B parameters active per token, built on Qwen3.5's hybrid of Gated DeltaNet linear attention and conventional attention. Its native window is 262,144 tokens; the card says it extends to about 1M with YaRN, and the hosted version ships with that longer window switched on.

What the hosted product adds is operational rather than architectural: the long window by default, built-in tools, and someone else's GPUs. What it takes away is control over the serving configuration and the ability to fine-tune. The Qwen3.5-397B-A17B guide covers the open model, its benchmark profile and what self-hosting it involves.

Two snapshots, and which one to pin

Alibaba lists two dated snapshots: qwen3.5-plus-2026-02-15, the launch build, and qwen3.5-plus-2026-04-20. On OpenRouter they appear as qwen/qwen3.5-plus-02-15 and qwen/qwen3.5-plus-20260420, with the later one listed at a slightly higher price.

The sources I could read do not publish a changelog for the April build, so do not assume it is strictly better at your task. A later snapshot of the same model usually carries post-training fixes, and those fixes can shift output format, verbosity and refusal behaviour as easily as they improve quality.

The safe pattern is to pin one snapshot in production, run the other against a fixed set of your own prompts, and switch only on evidence. The undated alias is convenient for experiments and a liability in production, because it can move under you.

The 256K price step

On Alibaba's international price list at the time of writing, Qwen3.5-Plus is tiered by the input size of each request:

• Up to 256K input tokens: $0.40 per million input, $2.40 per million output.
• 256K to 1M input tokens: $0.50 per million input, $3.00 per million output.

The important detail is that the tier applies to the whole request. Cross the line and every token, input and output, is billed at the higher rate. A request with 250K input and 4K output costs about $0.110. The same request at 260K input costs about $0.142, roughly 30% more for 4% more input.

That step is gentle compared with some providers, but it matters for agents whose transcripts drift past 256K late in a run. Compacting context before the threshold is often cheaper than paying the higher tier on every subsequent turn. Context window cost trade-offs covers the general version of this problem.

Third-party routers set their own prices for Plus, and at the time of writing OpenRouter's listing is below Alibaba's. Check the rate on the endpoint you actually call.

Plus against Flash in the same family

Qwen3.5-Flash is the hosted 35B-A3B, about 3B active against Plus's 17B. Alibaba prices it at $0.10 / $0.40 with no step across the 1M window. Below 256K, Plus costs four times as much for input and six times as much for output.

On the open models' own cards, the 397B leads the 35B-A3B on most reported measures: SWE-bench Verified 76.4 against 69.2, LiveCodeBench v6 83.6 against 74.6, MMMU 85.0 against 81.4. The gap is real but not dramatic, between about four and nine points on these three measures.

A reasonable split: Plus for harder reasoning, longer code changes and difficult visual questions; Flash for extraction, summarisation and anything high-volume. The Qwen3.5-Flash guide covers the cheaper side, and routing between them is usually better than choosing one.

Should you move to Qwen3.6-Plus

Alibaba also lists qwen3.6-plus, at $0.50 / $3.00 up to 256K and $2.00 / $6.00 above it. Below the threshold that is a modest increase over 3.5 Plus. Above it, input costs four times as much.

The Qwen3.6 open models report stronger agentic coding than their 3.5 counterparts, so 3.6 Plus is the one to test if your workload is code-heavy and prompts are mostly short. For long-context work, 3.5 Plus is the cheaper choice by a wide margin. The Qwen3.6-Plus guide has the details.

Checklist before you ship on Plus

1. Pin a dated snapshot, not the alias.
2. Log input tokens per request and alert when requests approach 256K.
3. Decide whether thinking is on for each route, and cap output tokens explicitly; the OpenRouter listing allows up to 65,536.
4. Run the same twenty prompts through Plus and Flash, and only keep Plus on routes where it wins clearly.
5. Re-run that set whenever you consider a newer snapshot or 3.6 Plus.