# Atlas with Qwen3-Next 80B-A3B Thinking: The Reasoning Tier in 2026 > Qwen3-Next 80B-A3B Thinking charges $0.50 per Mtok input and $6.00 per Mtok output, with a 128K tokens (131,072) context and a 32,768 token output cap. Qwen3-Next 80B-A3B Thinking is Alibaba's reasoning model built on the same 80B total and 3B active sparse body as Qwen3-Next Instruct, but trained to emit a reasoning trace before it answers. Inside Atlas it belongs on planning and hard debugging. The input price is identical to the Instruct variant at $0.50 per Mtok, and you pay for the deliberation on the output side, at $6.00 per Mtok, across a 128K tokens (131,072) context window. ## Key takeaways - Qwen3-Next 80B-A3B Thinking charges $0.50 per Mtok input, identical to the Instruct variant, so a large Atlas context costs the same whether you want thinking or not. - Output is $6.00 per Mtok, 3x the Instruct variant, and a reasoning model spends a lot of output tokens on the trace. - The model activates only about 3B of its 80B parameters per token, so deliberation is cheap on the input side. - Context is 131,072 tokens with 32,768 tokens of output shared between trace and answer, so a long trace can crowd out a long diff. - Pair it with `"small_model": "alibaba/qwen3-next-80b-a3b-instruct"` in atlas.json to keep background traffic off the $6.00 tier. ## Why run Atlas's plan agent on Qwen3-Next 80B-A3B Thinking? Atlas drafts a plan in a read-only plan agent and asks before switching to a build agent, and Qwen3-Next 80B-A3B Thinking is built for exactly that read-only phase. The model emits a reasoning trace before answering, so at $0.50 per Mtok input the deliberation reads the whole codebase cheaply. The plan agent in Atlas cannot write files. It reads, it searches, it argues with itself, and it produces a plan you approve before a single byte changes. Qwen3-Next 80B-A3B Thinking is the model to spend on there, because deliberation is what you are actually buying. Concentrate the reasoning budget on the plan, then switch to a cheaper model for the build. The economics work because Qwen3-Next 80B-A3B Thinking activates only about 3B of its 80B parameters per token, so the model deliberates at small-model compute cost on the input side. Atlas searches code with hybrid semantic and keyword retrieval fused by reciprocal rank fusion, which means the reasoning trace is grounded in the declarations that matter, not in files the model guessed at. ## What does Qwen3-Next 80B-A3B Thinking cost compared to the Instruct variant? Qwen3-Next 80B-A3B Thinking costs $0.50 per Mtok input and $6.00 per Mtok output. The input price is identical to the Instruct variant, meaning a large Atlas context costs the same whether you want thinking or not, but the $6.00 output tier is 3x the Instruct variant's $2.00. The pricing split on Qwen3-Next 80B-A3B Thinking tells you exactly where to spend. Input at $0.50 per Mtok is free of any thinking premium, so loading a big retrieved context into Qwen3-Next 80B-A3B Thinking costs no more than loading it into the Instruct model. Output at $6.00 per Mtok is where the trace lands, and a reasoning model spends a lot of output tokens on the trace before it ever reaches the answer. In an agentic loop that generates output on every turn, that adds up fast. The countermeasure inside Atlas is the small_model slot: set `"small_model": "alibaba/qwen3-next-80b-a3b-instruct"` so routine summaries and session titles skip the $6.00 output tier entirely and land on the $2.00 one. ## Does the 32,768 token output cap limit Qwen3-Next 80B-A3B Thinking? Yes. Qwen3-Next 80B-A3B Thinking gives you 131,072 tokens of context and 32,768 tokens of output shared between trace and answer. A long reasoning trace and a long unified diff compete for the same 32,768 tokens, and a verbose trace can crowd out the diff. The 32,768 token output cap on Qwen3-Next 80B-A3B Thinking is the constraint that catches people. Unlike a direct-answer model, where the entire output budget goes to the answer, a reasoning model spends part of that 32,768 on deliberation. Atlas computes a unified diff for every file edit and surfaces it for approval before writing, and a large multi-file diff is not small. If the trace runs long, the diff gets squeezed. The practical fix is structural, not prompt tuning: use Qwen3-Next 80B-A3B Thinking for the plan, where the output you want IS the reasoning, then hand the resulting plan to a model whose whole 32,768 tokens go to the edit. The 128K tokens (131,072) input window is not the bottleneck here; the output ceiling is. ## When is Qwen3-Next 80B-A3B Thinking the wrong choice? Qwen3-Next 80B-A3B Thinking is the wrong choice for mechanical work. At $6.00 per Mtok output, 3x the Instruct variant, paying for a reasoning trace on a rename or a boilerplate test file is money spent on deliberation you will never read. Route away from Qwen3-Next 80B-A3B Thinking whenever the task has a known shape. Renames, config edits, test scaffolding, and straightforward fixes do not benefit from a trace, and the $6.00 per Mtok output price makes them expensive. The Instruct sibling at $2.00 per Mtok output does the same job on the same 131,072 token context. The second case for switching away is output-heavy work that needs every token of the 32,768 ceiling for the artifact itself. And a third: if you need a repository-wide window, 128K tokens (131,072) will feel tight, and a larger-context Qwen tier is the better home. Atlas lets you switch the active model and provider on the fly with favorites and recents, so none of these are commitments. ## How do you configure Atlas to run Qwen3-Next 80B-A3B Thinking? Export DASHSCOPE_API_KEY, since Atlas reaches Alibaba through the OpenAI-compatible SDK, then set `"model": "alibaba/qwen3-next-80b-a3b-thinking"` in atlas.json. Verify the model resolves with `atlas models alibaba | grep thinking` before your first planning-heavy session in 2026. Configuration for Qwen3-Next 80B-A3B Thinking is a two-slot decision in atlas.json, not one. The main `"model"` goes to `alibaba/qwen3-next-80b-a3b-thinking` for planning-heavy sessions, and `"small_model"` goes to `alibaba/qwen3-next-80b-a3b-instruct` so routine summaries skip the $6.00 output tier. Both variants share the same provider and the same DASHSCOPE_API_KEY, so there is one credential to manage. Every Atlas tool call is permission-gated against allow, ask, and deny rules before it runs, which pairs well with a reasoning model: you read the trace, you read the plan, and you still approve each tool call. Atlas snapshots file changes as git patches so edits can be diffed and rolled back, so even a confidently-reasoned wrong answer from Qwen3-Next 80B-A3B Thinking is recoverable. ## Setup 1. Export DASHSCOPE_API_KEY in your shell, since Atlas reaches Alibaba through the OpenAI-compatible SDK. 2. Verify the model resolves with `atlas models alibaba | grep thinking`. 3. Set `"model": "alibaba/qwen3-next-80b-a3b-thinking"` in atlas.json for planning-heavy sessions. 4. Set `"small_model": "alibaba/qwen3-next-80b-a3b-instruct"` so routine summaries skip the $6.00 output tier. 5. Use Atlas's plan agent to concentrate the reasoning budget on the plan, then switch to a cheaper model for the build. 6. Watch the 32,768 token output ceiling on long refactors, since trace and diff share that budget. ## FAQ ### how much does qwen3-next 80b-a3b thinking cost per million tokens Qwen3-Next 80B-A3B Thinking costs $0.50 per Mtok input and $6.00 per Mtok output. The input price matches the Instruct variant exactly; the output price is 3x the Instruct variant's $2.00 per Mtok. ### what is the context window and output limit of qwen3-next 80b-a3b thinking Qwen3-Next 80B-A3B Thinking has a 128K tokens (131,072) context window and 32,768 tokens of output, which is shared between the reasoning trace and the answer. ### how do i set up qwen3-next 80b-a3b thinking in atlas Export DASHSCOPE_API_KEY, verify with `atlas models alibaba | grep thinking`, then set `"model": "alibaba/qwen3-next-80b-a3b-thinking"` in atlas.json. Atlas reaches Alibaba through the OpenAI-compatible SDK. ### is a reasoning model worth $6.00 per mtok output for coding For planning and hard debugging in Atlas, yes: Qwen3-Next 80B-A3B Thinking earns the $6.00 per Mtok output price in the read-only plan agent. For mechanical edits it does not, and the Instruct variant at $2.00 per Mtok output is the better fit. ### how do i keep reasoning model costs down in a coding agent Set a cheap small_model. In Atlas, `"small_model": "alibaba/qwen3-next-80b-a3b-instruct"` keeps routine summaries off the $6.00 per Mtok output tier, and running the plan agent first concentrates the reasoning budget where it pays. ### does the reasoning trace eat into the output budget Yes. Qwen3-Next 80B-A3B Thinking shares its 32,768 token output ceiling between the reasoning trace and the answer, so a long trace can crowd out a long unified diff. ### qwen3-next thinking vs instruct for atlas Qwen3-Next 80B-A3B Thinking emits a reasoning trace and costs $6.00 per Mtok output; Qwen3-Next 80B-A3B Instruct answers directly at $2.00 per Mtok output. Both share the same 80B total / 3B active body, the same $0.50 per Mtok input, and the same 131,072 token context. --- Canonical HTML: https://runatlas.sh/resources/models/qwen3-next-80b-a3b-thinking Source of truth: aeo_pages row `/resources/models/qwen3-next-80b-a3b-thinking` (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.