How much does Qwen3.8-Max-0902 cost?

$2 per million input tokens and $6 per million output tokens on QwenCloud — identical to the base Qwen3.8-Max. Cache hits are quoted at $0.17 for explicit caching and $0.25 for implicit caching, the first time the Max tier had been priced on QwenCloud, the API Alibaba introduced alongside Qwen3.8-Flash. A 30,000-in / 5,000-out task costs about $0.09.

Should I switch from Qwen3.8-Max to the 0902 snapshot?

If you write agentic code or office-automation workflows, yes — that is where the post-training gains land, at no price change. If you pin model strings, note that qwen3.8-max-0902 is separately addressable from the qwen3.8-max alias, so pointing at the alias means you already received this change on September 2, 2026 without a deploy.

Is Qwen3.8-Max-0902 competitive with frontier models?

Not on aggregate quality. Qwen3.8-Max sits near 57.7 on the Artificial Analysis Intelligence Index against Claude Fable 5.1 at 65.6, Claude Opus 5 at 63.0 and GPT-6 Astra at 61.1. Its case is price and context: a 1M-token window and a 131,072-token max output at $2/$6, which is roughly a fifth of GPT-6 Astra's input price.

Quick Answer

What Is Qwen3.8-Max-0902? Alibaba's Coding Snapshot

Q: What is Qwen3.8-Max-0902?

Qwen3.8-Max-0902 is a dated post-training snapshot of Alibaba's Qwen3.8-Max, released September 2, 2026. The architecture is unchanged — 2.4 trillion parameters, 1M-token context, 131,072 max output tokens — and so is the price at $2 per million input and $6 per million output. What changed is the post-training, which Alibaba's Qwen team targets at coding and what it calls Cowork: agentic software tasks and multi-step office work.

Published: September 6, 2026

The Short Answer

Qwen3.8-Max-0902 is a post-training snapshot of Alibaba’s Qwen3.8-Max, released September 2, 2026. Same 2.4-trillion-parameter architecture, same 1M-token context, same $2 / $6 price. The only thing that changed is what the model was trained to be good at — Alibaba’s Qwen team describes the target as “Coding & Cowork.”

If you were expecting a new frontier model, this is not one. If you run agents on Qwen, it is a free upgrade.

What Actually Changed

	Qwen3.8-Max	Qwen3.8-Max-0902
Parameters	2.4T	2.4T (unchanged)
Context	1,000,000	1,000,000 (unchanged)
Max output	131,072	131,072 (unchanged)
Input price	$2 / MTok	$2 / MTok (unchanged)
Output price	$6 / MTok	$6 / MTok (unchanged)
Post-training	general	coding + agentic “Cowork”
Released	earlier 2026	September 2, 2026

The gains show up on agentic coding and office-task benchmarks rather than on raw reasoning or math. That is the honest read: a checkpoint refresh, not a generation.

Why the Date in the Name

0902 is a dated snapshot, Alibaba following the same convention OpenAI popularised with gpt-4-0613. It matters operationally:

qwen3.8-max-0902 — pinned. Frozen behaviour, safe for regression-tested pipelines.
qwen3.8-max — alias. Tracks whatever Alibaba considers current, which means you received this change on September 2, 2026 with no deploy on your side.

If your evals moved on September 2 and nobody shipped anything, this is your explanation. Pin the snapshot if reproducibility matters more than freshness.

Pricing and the Caching Quirk

$2 input / $6 output per million tokens, unchanged from the base model.

Cache pricing is where it gets interesting: $0.17 for explicit cache hits, $0.25 for implicit. Explicit caching being cheaper than implicit is backwards from the usual arrangement — most vendors charge a premium for the control. Alibaba is effectively paying you to manage your own cache keys rather than letting the platform guess. For a RAG or long-system-prompt workload that reuses the same prefix thousands of times a day, taking explicit control is worth roughly a 32% discount on cached reads.

A reference task of 30,000 input and 5,000 output tokens costs about $0.09 — against $0.055 for Claude Haiku 4.5, $0.11 for Claude Sonnet 5 and $0.55 for GPT-6 Astra.

Where It Sits in September 2026

On the Artificial Analysis Intelligence Index as of September 2026:

Model	Index
Claude Fable 5.1	65.6
Claude Opus 5	63.0
Muse Spark 1.3	~62.0
GPT-6 Astra (max)	61.1
Kimi K3	59.6
GLM 5.3	59.4
Gemini 3.8 Flash	58.7
Qwen3.8-Max	57.7

It is the bottom of that list. That is not a dismissal — every model there costs more per token except the flash tier — but it does mean Qwen3.8-Max-0902 is a price-and-context play, not a capability play. Reach for it when a 1M-token window and a 131,072-token output ceiling at $2/$6 solve your problem, not when you need the smartest available answer.

The Context and Output Ceiling

Two specs deserve attention because they are unusual at this price:

1M-token context. Common at the frontier now, still rare at $2/MTok. At list price, filling the entire window once costs $2.00 — cheap enough to make whole-repository or whole-corpus prompting a realistic default rather than an occasional experiment.

131,072-token max output. This is larger than GPT-6 Astra’s 128,000 and double Gemini 3.8 Flash’s 64K. If your workload generates very long artifacts — full migrations, generated test suites, bulk translation — the output ceiling is often the binding constraint, and this is one of the highest available.

What to Watch Out For

⚠️ Informal announcement. The 0902 release went out via the Qwen team’s X account rather than a formal model card. Benchmark tables differ between third-party trackers as a result. Cite the source you personally measured against, and treat any single tracker’s numbers as provisional.

⚠️ Data residency. QwenCloud endpoints are Chinese-hosted. For many EU and US enterprises this is a hard block regardless of price or capability — check before you build.

⚠️ Open weights. Alibaba has released open weights across parts of the Qwen line, but do not assume the Max tier follows the same pattern or timing. Verify the license for the specific checkpoint before planning self-hosted deployment.

⚠️ Not a reasoning upgrade. The post-training targets coding and agentic office work. If your bottleneck is hard math or long-chain reasoning, this snapshot will not move your numbers.

Should You Use It?

Yes, if: you already run Qwen in production, you need 1M context or >128K outputs on a mid-tier budget, your workload is agentic coding, and Chinese-hosted inference is acceptable under your policy.

No, if: you need the highest available answer quality, image or audio input, US/EU data residency, or a formally documented model card for a compliance review.

Last verified: September 6, 2026.