Alibaba has released Qwen3.8-Max-0902, an updated snapshot of its flagship Qwen3.8-Max model. It keeps the 2.4-trillion-parameter base and the 1-million-token context window, but has been further post-trained specifically for coding and collaborative agent work. API pricing is unchanged at 2 USD (about 320 yen) per million input tokens and 6 USD (about 960 yen) per million output tokens, and the model has taken first place on the Code Arena: WebDev leaderboard for front-end development with 1,691 points.

An In-Place Upgrade One Month After the Initial Release

Qwen3.8-Max only became generally available in August, and the 0902 snapshot replaces its internals while keeping the same name and price. The API model ID is qwen3.8-max-0902, with a dated alias, qwen3.8-max-2026-09-02, also available. Pinning the dated ID means later snapshots will not silently change what your requests run against.

Alibaba describes the update as additional post-training for "Coding & Cowork." In other words, rather than changing the model's size or architecture, the team retrained it with a focus on writing code and on seeing tasks through to completion while juggling multiple tools.

What Improved

The official model page groups the improvements into three areas.

For coding, the emphasis is on engineering-scale projects that touch an entire codebase, and on long-horizon autonomous development that runs without a person watching every step, rather than one-off function writing.

For collaborative agents, Alibaba says the model shows greater composure when orchestrating multiple tools and is better at delivering tasks end to end.

Vision understanding was also refined, with sharper and more reliable results in chart reasoning, document parsing, and general visual perception. The model accepts image, text, and video input and produces text output.

Looking at the comparison table Alibaba published, the biggest gains appear in areas where the previous version was weak. TerminalBench 3.0, which measures agentic work in a terminal, rose from 11.3 to 29.0, and ProgramBench, which tests black-box program replication, went from 10.5 to 28.0, both more than doubling. JobBench, covering professional job tasks, climbed about 10 points from 53.4 to 64.0.

Set against Anthropic's Claude Opus 5 in the same table, however, Opus 5 still leads on most of the agentic coding and office-work rows. The 0902 snapshot comes out ahead on SWE-Atlas QnA (66.3), which tests repository-level code understanding, on Automation Bench (50.8) for agentic SaaS workflows, and on some visual-reasoning benchmarks. It is worth noting that several benchmarks in the table are Alibaba's own, and that scores for rival models were gathered from published figures.

First on Code Arena, Price Unchanged

In external evaluation, the 0902 snapshot scored 1,691 points and debuted at number one overall on Code Arena: WebDev, where human raters compare model outputs on real web-development tasks. According to Arena, which runs the leaderboard, that puts it 3 points above Claude Opus 5 (Max), 17 points above Kimi K3 (Max), and 22 points above the previous Qwen3.8-Max.

Despite the size of the update, API pricing has not moved. On QwenCloud, input costs 2 USD (about 320 yen) per million tokens and output 6 USD (about 960 yen). Input that hits the implicit cache is 0.25 USD (about 40 yen), explicit cache reads are 0.17 USD (about 27 yen), and explicit cache creation is 2.5 USD (about 400 yen). For long-running agents that re-read the same repository or document set every turn, these cache rates matter more to the real bill than the headline price.

※1 USD = 160 JPY (based on the September 1, 2026 closing rate)

The limits carry over from the previous snapshot: a maximum input of 991,000 tokens (983,000 with thinking mode), a maximum output of 131,000 tokens, up to 262,000 reasoning tokens, and a 1-million-token context window. Rate limits are 1 million tokens and 15,000 requests per minute. Built-in tools such as the code interpreter, web search, web extractor, and image search remain available, as do thinking mode, function calling, structured outputs, batch processing, and fine-tuning.

Switching Is Just a Model ID Change

For existing Qwen3.8-Max users, migration takes almost no effort. The QwenCloud API is offered in an OpenAI-compatible format, so changing the model name to qwen3.8-max-0902 is enough. Thinking mode can be enabled with an extra parameter.

from openai import OpenAI
import os

client = OpenAI(
    api_key=os.getenv("DASHSCOPE_API_KEY"),
    base_url="https://dashscope-intl.aliyuncs.com/compatible-mode/v1",
)

completion = client.chat.completions.create(
    model="qwen3.8-max-0902",
    messages=[{"role": "user", "content": "Who are you"}],
    extra_body={"enable_thinking": True},
    stream=True,
)

Other routes are already opening up beyond QwenCloud: Vercel's AI Gateway now offers the model as alibaba/qwen3.8-max-0902. Using Vercel's coding-agent setup, it can also be called from tools such as Claude Code and Codex.

Open Questions

A few things were left unsaid. One is open weights: the 0902 snapshot is only described as available through the QwenCloud API, and there is no word on whether it can be downloaded and run locally. Another is that Alibaba's comparison table lines up against Anthropic's Claude Fable 5 and Claude Opus 5, and does not include Claude Fable 5.1, which arrived at almost the same time. Those caveats are worth keeping in mind when quoting the numbers.

Summary

Qwen3.8-Max-0902 is a snapshot that retrains Alibaba's top model for coding and collaborative agent use. With 2.4 trillion parameters, a 1-million-token context, and pricing of 2 USD input and 6 USD output all unchanged, the agentic coding metrics that were weak in the previous version improved sharply, and the model took first place on Code Arena: WebDev. Existing users can migrate by changing a single model ID, though the gap to the top closed models and the question of open weights remain to be weighed.