Alibaba Qwen3.8-Max-0902 explained: benchmarks, pricing, context window, coding, reasoning, and API access
Alibaba released Qwen3.8-Max-0902 on September 2, 2026 as an upgraded snapshot of Qwen3.8-Max, with most of the improvement concentrated in software engineering, long-running autonomous work, tool orchestration and professional tasks.
The model keeps the 1 million-token context window of Qwen3.8-Max and supports both thinking and non-thinking operation.
It also accepts text, images and video, supports function calling and Alibaba's built-in tool ecosystem, and is available through Alibaba Cloud Model Studio under the qwen3.8-max-0902 model ID.
The update is particularly notable because Alibaba did not accompany the stronger coding performance with a higher international API list price.
At current international Model Studio pricing, Qwen3.8-Max-0902 costs $2 per million input tokens and $6 per million output tokens.
........
Specification | Qwen3.8-Max-0902 |
Release | September 2, 2026 |
Model ID | qwen3.8-max-0902 |
Alias | qwen3.8-max-2026-09-02 |
Context window | 1 million tokens |
Input | Text, images, video |
Output | Text |
Thinking mode | Supported |
Non-thinking mode | Supported |
Function calling | Supported |
Built-in tools | Supported |
Structured output | Supported |
International input price | $2 / 1M tokens |
International output price | $6 / 1M tokens |
Maximum images | 2,048 |
Maximum videos | 64 |
Maximum video duration | 2 hours |
API platform | Alibaba Cloud Model Studio |
........
The 0902 suffix identifies a dated Qwen3.8-Max snapshot, rather than an entirely separate model family.
Its value comes from what changed relative to the earlier Qwen3.8-Max rather than from a completely new product architecture.
··········
QWEN3.8-MAX-0902 TARGETS THE WEAK POINTS OF THE ORIGINAL RELEASE
The largest reported improvements are concentrated in coding and long-horizon execution rather than basic multimodal capability.
Alibaba describes the September snapshot as substantially stronger at engineering-scale software projects and autonomous development that continues across long sequences of actions.
The model is also designed to coordinate multiple tools during a task rather than treating each tool call as an isolated operation.
This becomes relevant when an AI system has to inspect a repository, identify dependencies, modify several files, run tests, diagnose failures and continue iterating without requiring a human to restart the workflow after every step.
The same principle extends beyond coding.
Alibaba reports improvements in professional tasks where the model must plan, use tools and deliver a completed result rather than simply answer a question.
The visual system was also refined for chart reasoning, document parsing and multimodal perception.
The practical change is concentrated in execution depth: Qwen3.8-Max-0902 is intended to remain useful after the initial prompt, when a task turns into a sequence of dependent decisions.
··········
CODING BENCHMARKS SHOW THE LARGEST GENERATIONAL GAINS
Alibaba's reported results show substantial improvements over the previous Qwen3.8-Max snapshot across repository-level and agentic software-engineering tests.
........
Coding benchmark | Qwen3.8-Max | Qwen3.8-Max-0902 | Change |
TerminalBench 3.0 | 11.3 | 29.0 | +17.7 |
DeepSWE 1.1 | 56.6 | 69.3 | +12.7 |
NL2Repo-Bench | 55.9 | 64.9 | +9.0 |
ProgramBench | 10.5 | 28.0 | +17.5 |
SWE-Marathon | 39.1 | 44.8 | +5.7 |
MLS-Bench-Lite | 41.0 | 50.1 | +9.1 |
SWE-Atlas QnA | 60.3 | 66.3 | +6.0 |
QwenSWEBench V2 | 55.1 | 70.0 | +14.9 |
........
TerminalBench rises from 11.3 to 29.0, approximately 2.57× the earlier score.
ProgramBench moves from 10.5 to 28.0, approximately 2.67× the previous result.
Those ratios should not be interpreted as the model becoming 2.6 times better at programming generally.
They show that the September post-training appears to have corrected specific weaknesses exposed by these evaluation environments.
DeepSWE increases from 56.6 to 69.3, while QwenSWEBench V2 moves from 55.1 to 70.0.
The benchmark package supports a narrower conclusion: the 0902 snapshot is materially stronger than the original Qwen3.8-Max on the software-engineering workloads Alibaba measured.
It does not establish that Qwen3.8-Max-0902 is universally the strongest coding model.
··········
THE MODEL IS CLOSER TO FRONTIER COMPETITORS BUT DOES NOT WIN EVERY CODING TEST
The benchmark profile is mixed enough to show where Qwen still has room to improve.
A model can perform well at repository understanding while remaining weaker at terminal execution, or perform well on long-horizon software work while losing on a different agent harness.
For production use, a benchmark measuring repository-level question answering does not test exactly the same behavior as an agent that must operate a terminal, edit files and recover from failed commands.
The most useful signal is therefore the breadth of improvement across multiple coding evaluations, rather than any individual leaderboard position.
Many of the currently available 0902 figures originate from Alibaba's own published comparison material and can depend on the harness, tool configuration, reasoning budget and execution environment used for each model.
Independent reproduction under identical conditions remains more informative than treating vendor benchmark tables as absolute rankings.
··········
LONG-HORIZON AGENT PERFORMANCE ALSO IMPROVED
The September update extends beyond code generation into professional and tool-driven work.
........
Agent/work benchmark | Qwen3.8-Max | Qwen3.8-Max-0902 | Change |
CoWorkBench | 74.8 | 76.1 | +1.3 |
JobBench | 53.4 | 64.0 | +10.6 |
Toolathlon Verified | 72.5 | 73.3 | +0.8 |
WorkArena Elo | 1,348 | 1,468 | +120 |
........
CoWorkBench and Toolathlon move only modestly, while JobBench gains 10.6 points and WorkArena rises by 120 Elo points in Alibaba's reported evaluation.
This indicates that the update did not produce a uniform capability increase: different behaviors improved by different amounts.
Agentic AI depends on several capabilities at once: understanding the request, decomposing the job, selecting tools, preserving state, checking intermediate results and recovering when an action fails.
Qwen3.8-Max-0902 is explicitly positioned around this longer execution chain.
··········
ONE MILLION TOKENS CHANGES THE TYPES OF WORKLOADS QWEN CAN ACCEPT
Qwen3.8-Max-0902 retains a 1 million-token context window.
That capacity makes it possible to place much larger working sets inside a single model interaction than with conventional 128K or 256K systems.
Possible workloads include large code repositories, collections of contracts, extensive financial documentation, research archives and long agent histories.
The model can also process up to 2,048 images or 64 videos, subject to platform limits, with supported video input extending to two hours and 2 GB.
A large advertised context window should not be interpreted as one million tokens of equally reliable recall.
Retrieval quality can vary with information position, prompt structure, task complexity and the amount of irrelevant material included in the context.
··········
THINKING AND NON-THINKING MODES SERVE DIFFERENT WORKLOADS
Qwen3.8-Max-0902 supports both reasoning-intensive and direct-response operation.
Thinking mode is intended for problems where additional internal computation can improve planning, mathematical reasoning, coding or multi-step execution.
Non-thinking operation is more appropriate when latency and predictable output matter more than extended reasoning.
A production application can reserve more expensive reasoning behavior for the minority of requests that actually require it.
··········
API PRICING IS AGGRESSIVE FOR A FLAGSHIP MODEL
Alibaba lists Qwen3.8-Max-0902 internationally at $2 per million input tokens and $6 per million output tokens.
The same international prices apply to the current Qwen3.8-Max listing.
The previous Qwen3.7-Max international price is listed at $2.50 per million input tokens and $7.50 per million output tokens.
........
Model | Input / 1M tokens | Output / 1M tokens |
Qwen3.7-Max | $2.50 | $7.50 |
Qwen3.8-Max-0902 | $2.00 | $6.00 |
Reduction | 20% | 20% |
........
Pricing varies by deployment scope, so the international rates should not be assumed to apply identically to every Alibaba Cloud region.
Context caching is also supported, creating another cost variable for applications that repeatedly send the same large prefix.
··········
A 1M-TOKEN CONTEXT WINDOW DOES NOT MEAN A $2 REQUEST
The API price is measured per million processed tokens, not per request with a one-million-token capacity.
A request containing 100,000 uncached input tokens at the $2-per-million rate costs approximately $0.20 for input.
If the model generates 10,000 output tokens, output costs approximately $0.06, for a combined token cost of roughly $0.26.
A workload containing 800,000 input tokens and 20,000 generated tokens would cost approximately $1.72.
At 100,000 such uncached executions, the token component alone would reach roughly $172,000.
This is why the 1M context window should be treated as capacity rather than a target.
··········
MULTIMODAL INPUT IS PART OF THE SAME AGENT WORKFLOW
Qwen3.8-Max-0902 is not limited to text and code.
The model accepts images and video natively and can combine visual information with function calling and built-in tools.
Alibaba specifically identifies improvements in chart reasoning, document parsing and multimodal perception for the September snapshot.
The model outputs text rather than generating images or video, so its multimodality is primarily understanding-oriented.
··········
BUILT-IN TOOLS REDUCE THE AMOUNT OF AGENT INFRASTRUCTURE DEVELOPERS MUST BUILD
Alibaba exposes web search, code execution and other tool capabilities through Model Studio for supported Qwen3.8 deployments.
Function calling is also available for applications that need their own tools or external systems.
Qwen3.8-Max-0902 is also included among the models supported through Alibaba Cloud's compatibility layer for the OpenAI Responses API.
Compatibility should not be confused with identical behavior: tool definitions, reasoning controls, token accounting, response objects and supported platform features still need to be validated before switching a production workload.
··········
THE 0902 SNAPSHOT IS MORE INTERESTING AS AN ENGINEERING UPDATE THAN AS A NEW NAME
Qwen3.8-Max-0902 demonstrates how much model behavior can change without changing the underlying product family.
The 1M context capacity remains, the multimodal interface remains, thinking and non-thinking modes remain, and the international $2/$6 token pricing remains.
What changes substantially is the measured ability to carry software-engineering and professional tasks through longer execution sequences.
Data Studios' calculation across the eight coding benchmarks for which comparable Qwen3.8-Max and 0902 figures are available gives an average absolute improvement of about 11.9 benchmark points.
That average should not be interpreted as a universal 11.9% capability increase because the benchmarks measure different tasks and use different scoring systems.
For an existing Qwen3.8-Max API user, qwen3.8-max-0902 is particularly relevant for repository-scale coding, autonomous development, professional agents and multi-tool workflows.
For a new user comparing frontier APIs, its strongest combination is 1M context, multimodal input, integrated tools and relatively low token pricing in a single flagship endpoint.
··········
FOLLOW US FOR MORE.
DATA STUDIOS
datastudios.org
