What are the strengths of this model?

Code Arena WebDev #1 at 1,691 — first Chinese model to top a blind, real-user-voted benchmark TerminalBench 3.0 more than doubled (11.3 → 29.0); QwenSWEbench V2 hit 70.0 OpenAI- and Anthropic-compatible APIs; direct integration with Claude Code, Codex, Qoder and Qwen Code Blended $5/MTok — best price-to-score ratio on the Code Arena Pareto frontier

What are the weaknesses of this model?

No open weights; API-only and routed through Chinese infrastructure under Chinese jurisdiction The 3-point WebDev lead over Opus 5 Max sits inside the confidence interval (±19 points) Still trails Opus 5 Max on TerminalBench, DeepSWE and agent-coordination benchmarks The '0902' suffix is a date stamp, not a version — behavior under the same model string can shift silently

What are the best use cases?

Front-end and repository-level coding where the Code Arena WebDev lead matters Long-horizon office automation (CoWorkBench-style workflows) Cost-sensitive code review and refactoring at $6 per million output tokens Drop-in OpenAI-compatible backend for Claude Code, Codex and Qoder

Alibaba Cloud (Qwen Team)Proprietary

Qwen3.8-Max-0902

Name: Qwen3.8-Max-0902
Author: Alibaba Cloud (Qwen Team)

Compare this model

Qwen3.8-Max-0902, deployed September 2, 2026 via QwenCloud, is a post-training refresh of Alibaba's Qwen3.8-Max. The underlying 2.4T-parameter, 95B-active MoE architecture is unchanged from the August 12 checkpoint, but the model climbed to #1 on Code Arena's blind WebDev leaderboard at 1,691 points and lifted TerminalBench 3.0 from 11.3 to 29.0. Context (1M tokens) and $2/$6 pricing are unchanged; availability is API-only across OpenAI/Anthropic-compatible endpoints in Beijing, Singapore and Virginia. No new open-weights drop — the downloadable Qwen3.8-Max remains the August 12 release.

Share:X Hatena LINE

Parameters

2.4T total / 95B active (MoE)

Context Window

1M (991K input / 131K output / 262K reasoning budget)

License

Closed (API-only via QwenCloud)

Release Date

2026-09-02

API Pricing

Input Price (per 1M tokens)

Output Price (per 1M tokens)

Billing Mode: blended $5/MTok, cache read $0.17

Strengths

•Code Arena WebDev #1 at 1,691 — first Chinese model to top a blind, real-user-voted benchmark
•TerminalBench 3.0 more than doubled (11.3 → 29.0); QwenSWEbench V2 hit 70.0
•OpenAI- and Anthropic-compatible APIs; direct integration with Claude Code, Codex, Qoder and Qwen Code
•Blended $5/MTok — best price-to-score ratio on the Code Arena Pareto frontier

Weaknesses

•No open weights; API-only and routed through Chinese infrastructure under Chinese jurisdiction
•The 3-point WebDev lead over Opus 5 Max sits inside the confidence interval (±19 points)
•Still trails Opus 5 Max on TerminalBench, DeepSWE and agent-coordination benchmarks
•The '0902' suffix is a date stamp, not a version — behavior under the same model string can shift silently

Use Cases

•Front-end and repository-level coding where the Code Arena WebDev lead matters
•Long-horizon office automation (CoWorkBench-style workflows)
•Cost-sensitive code review and refactoring at $6 per million output tokens
•Drop-in OpenAI-compatible backend for Claude Code, Codex and Qoder

Deep Analysis

Context Window

1M tokens

991K max input; 131K max output; 262K reasoning budget; text/image/video input

API Price

$2 / $6 per 1M

Input/output per million tokens; blended $5/MTok; cache read $0.17

Code Arena WebDev

#1 at 1,691

First Chinese model to top the blind leaderboard; 3 above Claude Opus 5 Max

TerminalBench 3.0

29.0

Up from 11.3 on the prior Qwen3.8-Max checkpoint (2.6x)

Parameters

2.4T total / 95B active

MoE; same architecture as Qwen3.8-Max, post-training refresh

License

Closed (API-only)

No new open-weights drop; checkpoint remains the Aug 12 release

Strengths

・Genuine, targeted coding lift: TerminalBench 3.0 more than doubled (11.3 to 29.0) and QwenSWEbench V2 hit 70.0, all at unchanged $2/$6 pricing.
・Top of Code Arena WebDev at 1,691 — the first Chinese model to lead that blind, real-user-voted benchmark.
・Best price-to-score on the leaderboard's Pareto frontier at a blended $5/MTok — roughly 1/5 the input cost of Fable 5.1.
・Drop-in integrations: OpenAI- and Anthropic-compatible APIs plus Claude Code, Codex, Qoder and Qwen Code clients.

Weaknesses

・Not an open-weights release — API-only on Chinese infrastructure, so every call answers to Chinese jurisdiction.
・The 3-point WebDev lead over Opus 5 Max is inside the confidence interval (plus/minus 19 points, ~1,389 votes vs Opus's 10,000+); don't over-read it.
・Still trails Opus 5 on most of Qwen's own comparison rows: TerminalBench, DeepSWE, CoWorkBench, Toolathlon.
・The 0902 is a date-stamped snapshot, not a version — behavior can shift under the same model string without notice; pin your snapshot.

Competitor Comparison

Model	Arena	SWE	GPQA	Price
Qwen3.8-Max-0902	#1 WebDev 1,691	70.0 (QwenSWE V2)	N/A	$2/$6
Claude Opus 5 Max	WebDev 1,688	70.0 (QwenSWE V2)	N/A	$10/$50
Kimi K3 Max	WebDev 1,674	N/A	N/A	N/A
GLM-5.3 Flash	57 (AA Index)	63.4 (Z.ai)	N/A	$0.15/$0.50

Overview

Qwen3.8-Max-0902, shipped September 2, 2026, is a post-training refresh of Alibaba's 2.4T-parameter flagship — same 1M context, same $2/$6 pricing — that climbed to #1 on Code Arena's WebDev leaderboard at 1,691 points. It is the first Chinese-developed model to top that blind, real-user-voted benchmark, and it did so with a 22-point gain over its own prior checkpoint. The story is not new architecture; it is that a large share of perceived model quality is now post-training, delivered at unchanged price.

Benchmarks & Performance

Alibaba reports TerminalBench 3.0 up from 11.3 to 29.0, ProgramBench (Almost Solved) 10.5 to 28.0, QwenSWEbench V2 55.1 to 70.0, DeepSWE 1.1 56.6 to 69.3, JobBench 53.4 to 64.0, and WorkArena Elo 1,348 to 1,468. On Code Arena WebDev it leads at 1,691 vs Claude Opus 5 Max 1,688 and Kimi K3 Max 1,674. Most numbers are Alibaba-reported and not yet independently replicated.

Detailed Comparison

Against Claude Opus 5 Max it wins on three coding rows (MLS-Bench-Lite, SWE-Atlas QnA, QwenSWEBench V2), WorkArena, and both multimodal evals, while trailing on TerminalBench, DeepSWE and agent-coordination benchmarks. Against GPT-6 Astra it is ~8x cheaper per output token and a better Pareto play for front-end and repository work, though not a replacement for the hardest agentic tasks. The 3-point WebDev margin is statistically thin.

Community Feedback

The 0902 suffix — a deployment-date stamp rather than a version bump — is how Chinese labs now ship capability without triggering new-model expectations. Developers relying on reproducible behavior were caught off guard: the unpinned qwen/qwen3.8-max slug began routing to the new snapshot on September 5. Coverage framed it as China's first front-end AI crown at one-fifth rival prices.

Use Cases

Best fit: front-end and repository-level coding, long-horizon office automation (CoWorkBench-style), and cost-sensitive agentic pipelines where $6/output-token beats $50. Route it as a cheap tier for code review and refactors, not as a drop-in for the hardest reasoning. Enterprises must settle the Chinese-jurisdiction question before any pilot, since the only access path is QwenCloud on Chinese infrastructure.

Latest News

Alibaba pushed Qwen3.8-Max-0902 live on September 2, 2026 via QwenCloud (id qwen3.8-max-0902) with OpenAI- and Anthropic-compatible endpoints in Beijing, Singapore and Virginia. It debuted #1 on Code Arena WebDev at 1,691. The update is post-training only — no new open-weights drop, and the downloadable checkpoint remains the August 12 release, so self-hosting does not get these numbers.

Sources

Analysis generated: 2026-09-06