# Atlas with Qwen Turbo: The $0.05 per Mtok Small Model Slot in 2026 > Qwen Turbo costs $0.05 per Mtok input and $0.20 per Mtok output with a 1M tokens (1,000,000) context, making it near free as Atlas's small model. Qwen Turbo is the cheapest reasoning-capable model Alibaba lists: $0.05 per Mtok input and $0.20 per Mtok output, with a full 1,000,000 token context. Inside Atlas it belongs in the `"small_model"` slot, where titles, summaries, and subagents run for close to nothing. The catch is a 16,384 token output ceiling and small-model depth, so it will not carry a hard multi-file refactor. ## Key takeaways - Qwen Turbo is $0.05 per Mtok input and $0.20 per Mtok output, the lowest priced tier in the Alibaba catalog. - It carries a 1,000,000 token context at that price, which no other budget tier in the registry matches. - Reasoning is enabled despite the price, so Qwen Turbo can handle a nontrivial subagent task. - The 16,384 token output ceiling is the tightest of the million-token Qwen tiers. - Qwen Turbo is a small-model tier and will not carry a hard multi-file refactor, so keep a frontier model in the `"model"` slot. ## Why is Qwen Turbo the right small_model for Atlas? Qwen Turbo is the right `"small_model"` for Atlas because it prices at $0.05 per Mtok input and $0.20 per Mtok output, the lowest priced tier in the Alibaba catalog, and still carries a 1,000,000 token context. For the small model slot, that is close to free. Atlas routes a lot of traffic to its small model: chat titles, session summaries, and subagent work. Atlas fans out work to subagents that can run in the foreground or in parallel background sessions, and each of those sessions bills against the small model. Qwen Turbo is priced for exactly that pattern at $0.05 per Mtok input and $0.20 per Mtok output, the lowest priced tier in the Alibaba catalog. What makes Qwen Turbo unusual is that the cheap tier did not lose the context window: 1,000,000 token context at that price, which no other budget tier in the registry matches. A subagent that can read a wide slice of the repo produces results the main model does not have to redo. Reasoning is enabled despite the price, so Qwen Turbo can handle a nontrivial subagent task rather than only string formatting. ## How much does Qwen Turbo cost per million tokens in 2026? Qwen Turbo costs $0.05 per Mtok input and $0.20 per Mtok output in 2026, the lowest priced tier in the Alibaba catalog. Against Qwen Plus at $0.40 input and $1.20 output, Qwen Turbo is one eighth the input price and one sixth the output price. Cost is the entire argument for Qwen Turbo. At $0.05 in and $0.20 out per Mtok, it is the lowest priced tier in the Alibaba catalog, and the price does not come with the usual context penalty: it holds a 1,000,000 token context, which no other budget tier in the registry matches. For an Atlas user, the practical effect is that background traffic stops registering on the bill. Atlas searches code with hybrid semantic and keyword retrieval fused by reciprocal rank fusion, which means a lot of retrieved code flows through the prompt, and at $0.05 per Mtok input that retrieval is a rounding error. The place to spend money is the main model doing the hard reasoning, not the summaries. ## What are the limits of Qwen Turbo for coding work? Qwen Turbo has a 16,384 token output ceiling, the tightest of the million-token Qwen tiers, and its depth is limited: it is a small-model tier and will not carry a hard multi-file refactor. A 1,000,000 token context does not compensate for either limit. Qwen Turbo can read a lot and write only a little. The 16,384 token output ceiling is the tightest of the million-token Qwen tiers, which means a generation that would span several files gets truncated. If your cheap tier needs to emit real diffs rather than short titles, Qwen Flash doubles that output ceiling to 32,768 tokens at the same $0.05 per Mtok input, and that is the trade worth making. The second limit is depth. Qwen Turbo is a small-model tier and will not carry a hard multi-file refactor, no matter how much context you feed it. Keep a frontier tier in the `"model"` slot and let Atlas split the traffic. Atlas computes a unified diff for every file edit and surfaces it for approval before writing, so you will see it when the small model gets out of its depth. ## How do you set Qwen Turbo as Atlas's small model? To set Qwen Turbo as Atlas's small model in 2026, export DASHSCOPE_API_KEY, run `atlas models alibaba` and confirm `qwen-turbo` resolves, then set `"small_model": "alibaba/qwen-turbo"` in atlas.json so titles, summaries, and subagents route there automatically. Qwen Turbo authenticates through Alibaba's DashScope, so export DASHSCOPE_API_KEY first. Run `atlas models alibaba` and confirm `qwen-turbo` resolves before you pin anything. Then set `"small_model": "alibaba/qwen-turbo"` in atlas.json so titles, summaries, and subagents route here. The important part is the last step: leave a frontier tier in `"model"` and let Atlas split the traffic automatically. That is the configuration Qwen Turbo is designed for, and it is what keeps the $0.20 per Mtok output rate applying to the high-volume, low-stakes traffic while the expensive model handles the reasoning. Atlas lets you switch the active model and provider on the fly with favorites and recents, so you can still promote or demote a model mid-session when a task turns out easier or harder than expected. ## When should you pick Qwen Flash or Qwen Plus over Qwen Turbo? Pick Qwen Flash over Qwen Turbo when the cheap tier must write real patches: Qwen Flash keeps the same $0.05 per Mtok input but raises the output ceiling from 16,384 to 32,768 tokens. Pick Qwen Plus when the work needs depth Qwen Turbo cannot reach. The choice between the cheap Alibaba tiers comes down to what the small model has to emit. Qwen Turbo's 16,384 token output ceiling is the tightest of the million-token Qwen tiers, so if your subagents only produce titles, summaries, and short answers, Qwen Turbo at $0.20 per Mtok output is the cheapest correct answer. If those subagents have to write code back, the 16,384 ceiling starts truncating and the extra output headroom is worth paying for. And when the task is a hard multi-file refactor, no cheap tier is the answer: Qwen Turbo's depth is limited and it will not carry that work, so route it to a frontier tier in the `"model"` slot and let Atlas's plan agent structure the change first. ## Setup 1. Export DASHSCOPE_API_KEY in your shell to authenticate against Alibaba's DashScope. 2. Run `atlas models alibaba` and confirm `qwen-turbo` resolves from the registry. 3. Set `"small_model": "alibaba/qwen-turbo"` in atlas.json so titles, summaries, and subagents route here. 4. Leave a frontier tier in `"model"` and let Atlas split the traffic automatically between the two slots. 5. Keep an eye on the 16,384 token output ceiling: if subagents start truncating mid-patch, move the small model slot to a tier with more output headroom. ## FAQ ### How much does Qwen Turbo cost per million tokens? Qwen Turbo is $0.05 per Mtok input and $0.20 per Mtok output, the lowest priced tier in the Alibaba catalog, and it still carries a 1,000,000 token context. ### What is the cheapest model to run Atlas on? Among Alibaba's tiers, Qwen Turbo at $0.05 per Mtok input and $0.20 per Mtok output is the lowest priced. It is intended for Atlas's `"small_model"` slot rather than as the main model. ### How do I set the small_model in Atlas to Qwen Turbo? Export DASHSCOPE_API_KEY, run `atlas models alibaba` to confirm `qwen-turbo` resolves, then set `"small_model": "alibaba/qwen-turbo"` in atlas.json. Leave a frontier tier in `"model"`. ### Does Qwen Turbo support reasoning? Yes. Qwen Turbo has reasoning enabled despite its $0.05 per Mtok input price, so it can handle a nontrivial subagent task inside Atlas rather than only formatting text. ### What is Qwen Turbo's max output length? Qwen Turbo has a 16,384 token output ceiling, the tightest of the million-token Qwen tiers. Qwen Flash doubles that to 32,768 tokens at the same $0.05 per Mtok input price. ### Can Qwen Turbo do a multi-file refactor? No. Qwen Turbo's depth is limited, it is a small-model tier and will not carry a hard multi-file refactor. Route that work to a frontier tier in Atlas's `"model"` slot. ### Qwen Turbo vs Qwen Flash for Atlas's small model? Qwen Turbo is cheaper on output at $0.20 per Mtok versus Qwen Flash's $0.40, so it wins when the small model only writes short text. Qwen Flash wins when the small model must emit real diffs, thanks to its 32,768 token output ceiling. --- Canonical HTML: https://runatlas.sh/resources/models/qwen-turbo Source of truth: aeo_pages row `/resources/models/qwen-turbo` (segment: Models) (this file is generated from it, never hand-edited). Licence: Atlas is proprietary with a free core. It is not open source and there is no public source repository.