Qwen3.8-Max is live at $2/$6 per million tokens — and Alibaba's own card shows Fable 5 winning all three SWE benchmarks

Alibaba released Qwen3.8-Max on August 3, 2026, three weeks after teasing it as a preview. (Source: Qwen official blog, retrieved 2026-08-06)

Key facts:

• The model is 2.4 trillion parameters with 95 billion active per token — a sparse MoE that routes about 4% of its weights per forward pass.
• The context window is 1 million tokens. Inputs are text, image and video; output is text only.
• Pricing is $2.00 per million input tokens and $6.00 per million output tokens. Implicit cache reads are $0.25/M; explicit cache writes are $2.50/M and explicit cache reads $0.17/M. (Source: Alibaba Cloud pricing via Pasquale Pillitteri, 2026-08-03)
• It is served through QwenCloud and Alibaba Cloud Model Studio, with OpenAI-compatible Chat Completions and Responses APIs plus an Anthropic-compatible endpoint, and reasoning effort settable to low, medium or xhigh. (Source: TestingCatalog, 2026-08-03)
• Open weights have not shipped. Alibaba says weights for Qwen3.8-Max plus a smaller Qwen3.8-27B checkpoint land on Hugging Face and ModelScope the week of August 10. As of this writing, neither is up.
• QwenWork, Alibaba’s agent workspace, entered public beta the same day on web and desktop. (Source: SCMP, 2026-08-03)

What this means if you’re building with Qwen3.8-Max

1. The “beats GPT-5.6 Sol and Fable 5” headline is a workload claim, not a ranking. On Alibaba’s own card, Claude Fable 5 wins all three software-engineering benchmarks — SWE-Pro 80.0 vs 67.7, FrontierSWE 88.8 vs 73.5, and it also leads QwenReactBench (1770 vs 1724), CoWorkBench (75.9 vs 74.8), JobBench (57.4 vs 53.4) and MobileWorld (85.5 vs 77.8). GPT-5.6 Sol takes TerminalBench-2.1 (88.8 vs 86.6). Where Qwen3.8-Max leads is research reproduction (PaperBench 93.0 vs Sol’s 90.5 and Fable’s 88.8), agentic computer use (OSWorld-Verified 86.1 vs 85.0), embodied and visual reasoning (ERQA 77.8, PerceptionBench 63.5) and long video (LVBench 81.8). If your workload is writing and fixing code, this card does not say to switch.

2. The most useful chart in the release is the one nobody quoted. Alibaba ran Qwen3.8-Max inside five different agent harnesses, not just its own.

That spread is the practical signal: the model does not collapse when you take it out of Alibaba’s own scaffolding, which is the failure mode that kills most “drop-in replacement” plans. JobBench is the same story — 57.5 to 59.8 across four third-party harnesses, against 31.3 for Qwen3.7-Max.

3. Price the cache, not the sticker. At $2.00/$6.00, the interesting number is the $0.25/M implicit cache read — an 8x discount on input for repeated prefixes. Agent loops that resend a long system prompt and tool schema every turn are exactly the shape that discount is built for. Budget with cache hits assumed, or you will over-forecast by a lot.

4. Do not plan around weights that have not shipped. This is the second dated open-weights promise for Qwen 3.8 in three weeks; the July 19 preview banner said “coming soon” too. If you need weights you can hold today, GLM-5.2 and Kimi K3 are shipped. The safe move is the same one that applied to the preview: keep base_url and model ID behind a config flag, since the API is OpenAI- and Anthropic-compatible, and swap when the weights and an independent benchmark both exist.

Sources: Qwen official blog · SCMP, 2026-08-03 · TestingCatalog, 2026-08-03 · OfficeChai, 2026-08-03 · Pasquale Pillitteri, 2026-08-03

Related: Qwen3.7-Max · How to access Qwen3.8-Max Preview · Qwen 3.6 for local coding