Key Takeaways
- Qwen3.8-Max is a 2.4-trillion-parameter mixture-of-experts model released August 3, 2026 by Alibaba’s Qwen team, with 95 billion active parameters per forward pass, a 1-million-token context window, and native text, image, and video input. It is Alibaba’s first Max-class model with a committed open-weight release, making it the only frontier-tier model in the 2026 class that teams will be able to deploy on their own infrastructure once the weights ship.
- Qwen3.8-Max scored 86.1 on OSWorld-Verified, the highest reported score among frontier models, ahead of Claude Fable 5 at 85.0, GPT-5.6 Sol Max at 83.2, and Gemini 3.1 Pro at 76.2. On PaperBench, which measures autonomous machine learning research replication, it scored 93.0, ahead of GPT-5.6 Sol at 90.5 and Claude Fable 5 at 88.8. These are the two benchmarks most directly measuring long-horizon autonomous task execution, which is the model’s stated design target.
- Alibaba demonstrated Qwen3.8-Max running a 16-day autonomous software engineering run on the oh-my-cli open-source project, ingesting community-reported issues: generating code fixes, running end-to-end tests, and merging pull requests on GitHub without human intervention across the full duration. This demonstration is the longest publicly documented autonomous coding run by a frontier model as of August 2026.
- On SWE-bench Pro, a benchmark measuring issue-level code repair, Qwen3.8-Max scores 67.7, behind Claude Fable 5 at 80.0 and Opus 4.8 at 69.2. This creates a clear profile: Qwen3.8-Max leads on long-horizon autonomous work and agentic task completion, but trails Anthropic’s models on the classic issue-fixing benchmarks that matter most for teams using AI to resolve individual GitHub issues at scale.
- On GPQA Diamond, Qwen3.8-Max scores 92.6, level with Claude Fable 5. On Humanity’s Last Exam, it scores 43.6%, trailing Claude Fable 5 at 53.3% and GPT-5.6 Sol at 47.2%. On Terminal-Bench 2.1, it scores 86.6. The benchmark pattern shows a model that is competitive or leading on agentic and environment-interaction tasks, competitive on scientific reasoning, and below best-in-class on classic text-based reasoning benchmarks.
- Pricing at QwenCloud is $2 per million input tokens, $6 per million output tokens, and $0.25 per million implicitly cached tokens. At $2/$6: it is directly price-competitive with Grok 4.5 and Claude Fable 5 on the input side, while the $0.25 cached token rate offers significant savings for agentic workflows that reuse large system prompts or context windows repeatedly across a long task chain.
- The model processes video natively, with Alibaba demonstrating the ability to ingest 100-plus-hour video streams and build continuous memory graphs from them. This is functionally unique among the models it competes with most directly (Claude Fable 5, GPT-5.6 Sol Max) at the API level. The use case is enterprise workflows where video is a primary information source, such as surveillance analysis, training data review, or long-form meeting and presentation summarization.
Qwen3.8-Max arrives at a moment when the frontier coding model race has effectively split into two categories: models that are best at fixing discrete issues and models that are best at running long autonomous tasks from start to finish. Most frontier models have been optimized for the first category, because SWE-bench and similar issue-level benchmarks have dominated the evaluation narrative. Qwen3.8-Max is explicitly designed for the second.
This review covers the full architecture, every major benchmark score in context, the 16-day autonomous coding demonstration, pricing at QwenCloud, the open-weight release timeline, and an honest comparison against Claude Fable 5, GPT-5.6 Sol Max, and Kimi K3 for the teams most likely to evaluate it.
What Is Qwen3.8-Max?
Qwen3.8-Max is Alibaba’s flagship AI model and the largest model the Qwen team has released to date. It uses a mixture-of-experts (MoE) architecture with 2.4 trillion total parameters and 95 billion active parameters per forward pass. The MoE design means the full parameter count is partitioned into specialized expert subnetworks, with only a subset activated for any given input, which allows the model to apply specialist depth to different task types while keeping inference cost lower than a dense model of equivalent total parameter size.
The model supports a 1-million-token context window and accepts text, image, and video as native inputs. It was previewed at the World AI Conference in Shanghai on July 19, 2026, and fully released on August 3, 2026. Alibaba has committed to open-sourcing the weights alongside a smaller Qwen3.8-27B dense model, making Qwen3.8-Max the first Max-class model from any major lab planned for public open-weight release in 2026.
Qwen3.8-Max Features
Autonomous Long-Horizon Coding
The headline capability Alibaba demonstrated at launch is a 16-day autonomous software engineering run on the oh-my-cli open-source project. The model operated without human intervention across the full duration: it read incoming community issues on GitHub, determined which issues were valid, wrote code to address them, executed end-to-end test suites, and merged pull requests when tests passed. Alibaba describes the run as the longest publicly documented autonomous coding demonstration by a frontier AI model.
The architectural reason for this capability is the combination of MoE specialist routing, long context retention across a 1-million-token window, and tool-use capability that allows the model to operate development environment tooling, including compilers, test runners, and version control systems, as a native capability rather than through external agent scaffolding. Qwen3.8-Max’s 86.6 on Terminal-Bench 2.1 and 86.1 on OSWorld-Verified reflect this environment-operation strength.
Native Multimodal Input Including Long Video
Qwen3.8-Max accepts text, images, and video natively. The video capability extends to 100-plus-hour streams, with the model building continuous memory graphs from the video content rather than processing it as a static document. This is meaningful for enterprise use cases where video is a primary information source: surveillance review, training and onboarding video analysis, long-form meeting documentation, and product usage recordings. The 1-million-token context window in combination with video processing means the model can maintain coherent understanding across extended video inputs that would overwhelm shorter-context alternatives.
PaperBench and Research Replication
On PaperBench, which tests a model’s ability to autonomously replicate machine learning research papers including implementing the experimental code and matching the original results, Qwen3.8-Max scores 93.0. This is the highest published score on the benchmark, ahead of GPT-5.6 Sol at 90.5 and Claude Fable 5 at 88.8. PaperBench directly measures the capability most relevant to research teams that need to verify published results, extend prior work, or run systematic comparisons across research implementations, and Qwen3.8-Max’s lead here is the clearest benchmark evidence of its autonomous research task advantage.
OSWorld-Verified Environment Interaction
OSWorld-Verified measures a model’s ability to operate computer interfaces to complete tasks, including file management, application operation, web browsing, and multi-application workflows. At 86.1, Qwen3.8-Max scores above Claude Fable 5 at 85.0, GPT-5.6 Sol Max at 83.2, and Gemini 3.1 Pro at 76.2. This benchmark directly reflects real-world agentic deployment quality where the model needs to understand and interact with software environments rather than only generate text.
Qwen3.8-Max Pricing
| Token type | Price per million tokens |
|---|---|
| Input tokens | $2.00 |
| Output tokens | $6.00 |
| Cached input tokens | $0.25 |
The API is available through QwenCloud at launch pricing. At $2.00 input and $6.00 output per million tokens, Qwen3.8-Max is priced in line with Grok 4.5 and at parity with Claude Fable 5 on the input side. The $0.25 cached token rate is the most important pricing feature for agentic workflows: a model running a 16-day coding task that repeatedly references a large repository context across thousands of calls would pay $0.25 per million tokens on the repeated context rather than $2.00, which reduces the effective cost of long-horizon tasks substantially. No free tier is available at launch. Open-weight deployment on self-hosted infrastructure will be possible once the weights release, eliminating per-token costs entirely for teams with sufficient compute.
Qwen3.8-Max Pros and Cons
Pros:
- Highest OSWorld-Verified score (86.1) of any frontier model as of August 2026, demonstrating leading real-world computer environment interaction
- Highest PaperBench score (93.0) for autonomous research replication, ahead of GPT-5.6 Sol and Claude Fable 5
- 16-day autonomous coding run is the longest public demonstration of sustained AI software engineering
- Native video input supporting 100-plus-hour streams with continuous memory graph construction
- Competitive $2/$6 per million token pricing with $0.25 cached token rate that significantly reduces long-horizon task costs
- First Max-class model with committed open-weight release, enabling self-hosted deployment once weights ship
- GPQA Diamond at 92.6 is competitive with Claude Fable 5, confirming strong scientific reasoning alongside agentic strengths
Cons:
- SWE-bench Pro at 67.7 trails Claude Fable 5 at 80.0 and Opus 4.8 at 69.2, meaning it is not the best choice for discrete issue-fixing at scale
- HLE at 43.6% trails Claude Fable 5 at 53.3% and GPT-5.6 Sol at 47.2% on expert-level question-answering
- Open weights not yet available at launch, limiting self-hosted deployment for now
- Benchmark claims from Alibaba have not been fully independently verified as of August 2026; OSWorld-Verified and PaperBench results are primarily vendor-reported
- No free tier; teams must pay QwenCloud rates to evaluate the model before the open-weight release
Qwen3.8-Max vs Alternatives
Qwen3.8-Max vs Claude Fable 5: Claude Fable 5 leads on SWE-bench Pro (80.0 vs 67.7), Humanity’s Last Exam (53.3% vs 43.6%), and the issue-level coding benchmarks that most engineering teams use to evaluate AI coding tools. Qwen3.8-Max leads on OSWorld-Verified (86.1 vs 85.0) and PaperBench (93.0 vs 88.8), the benchmarks most directly measuring agentic task completion and autonomous research replication. The choice between them maps cleanly to task type: Claude Fable 5 for issue resolution pipelines, Qwen3.8-Max for long-horizon autonomous workflows and research tasks. They are priced comparably at $2/$6 per MTok.
Qwen3.8-Max vs GPT-5.6 Sol Max: GPT-5.6 Sol Max scores 83.2 on OSWorld-Verified (below Qwen’s 86.1), 90.5 on PaperBench (below Qwen’s 93.0), and 47.2% on HLE (above Qwen’s 43.6%). On SWE-bench Pro, GPT-5.6 Sol scores 64.6 compared to Qwen3.8-Max at 67.7, making it the one major coding benchmark where Qwen holds a narrow lead. GPT-5.6 Sol Max is a stronger general-purpose reasoning model; Qwen3.8-Max is the stronger autonomous agent. For teams already deployed on OpenAI’s API, the switching cost to QwenCloud needs to be weighed against Qwen’s agentic benchmark lead.
Qwen3.8-Max vs Kimi K3: Kimi K3, released by Moonshot AI earlier in 2026, targets the same long-context and agentic task market as Qwen3.8-Max. Qwen3.8-Max surpasses Kimi K3 on PaperBench and OSWorld-Verified, while Kimi K3 is more cost-competitive on standard inference pricing. Teams evaluating either model for research agentic use cases should run both on their specific task type rather than relying on aggregate benchmarks, as performance differences vary significantly by domain.
Who Is Qwen3.8-Max Best For?
Qwen3.8-Max is best suited for four use cases where its architectural strengths translate directly to production value. First, engineering teams building autonomous software agents that run multi-day task pipelines without human supervision. The 16-day coding demonstration and 86.6 Terminal-Bench score reflect a model built to sustain reliability across extended agentic runs rather than optimized only for individual inference quality.
Second, research teams that need to replicate, extend, or compare machine learning research at scale. A PaperBench score of 93.0 is directly applicable to automated literature verification, systematic experiment reproduction, and research-driven product development where reviewing and building on prior work is a routine workflow step.
Third, enterprises processing video as a primary information source. The 100-plus-hour video processing capability with memory graph construction is not matched at the API level by Claude Fable 5 or GPT-5.6 Sol Max, making Qwen3.8-Max the practical choice for video-heavy enterprise workflows until competitors close that gap.
Qwen3.8-Max is less well suited for teams primarily evaluating AI on discrete issue-fixing benchmarks (SWE-bench Pro) or expert-level question-answering (HLE), where Claude Fable 5 holds a meaningful lead. Teams running high-volume issue triage and resolution pipelines should evaluate Claude Fable 5 first and consider Qwen3.8-Max for the specific autonomous workflow scenarios where it leads.
Our Verdict
Qwen3.8-Max is the most capable autonomous AI agent available through a commercial API as of August 2026, on the benchmarks that directly measure agentic task execution: OSWorld-Verified (86.1), PaperBench (93.0), and Terminal-Bench 2.1 (86.6). The 16-day autonomous coding run is the most compelling real-world demonstration of sustained agentic capability published by any frontier lab. The $2/$6 pricing is competitive, the cached token rate reduces long-horizon task economics, and the planned open-weight release will make it the first Max-class model deployable on self-hosted infrastructure. The honest caveat is that its SWE-bench Pro score and HLE score trail Claude Fable 5, so teams that evaluate on issue-level coding benchmarks should test both models on their actual workloads rather than assuming the agentic benchmark lead translates to their specific use case.
Frequently Asked Questions
What is Qwen3.8-Max?
Qwen3.8-Max is Alibaba’s flagship AI model released August 3, 2026. It is a 2.4-trillion-parameter mixture-of-experts model with 95 billion active parameters, a 1-million-token context window, and native text, image, and video input. It is designed for autonomous long-horizon coding tasks, professional research work, and agentic workflows that run continuously without human supervision. The API is available through QwenCloud at $2 per million input tokens and $6 per million output tokens.
How does Qwen3.8-Max benchmark against Claude Fable 5?
Qwen3.8-Max leads Claude Fable 5 on OSWorld-Verified (86.1 vs 85.0) and PaperBench (93.0 vs 88.8), the benchmarks most directly measuring agentic task completion and autonomous research replication. Claude Fable 5 leads on SWE-bench Pro (80.0 vs 67.7), Humanity’s Last Exam (53.3% vs 43.6%), and most issue-level coding benchmarks. Both models score 92.6 on GPQA Diamond. The choice depends on task type: Qwen for sustained autonomous workflows, Claude Fable 5 for issue-level coding and expert reasoning.
What is the 16-day autonomous coding run?
Alibaba demonstrated Qwen3.8-Max running autonomously for 16 consecutive days on the oh-my-cli open-source GitHub project. During the run, the model read incoming community-reported issues, determined which were valid, wrote code to address them, executed end-to-end test suites, and merged pull requests when tests passed. No human intervention occurred during the 16-day period. Alibaba describes this as the longest publicly documented autonomous software engineering run by a frontier AI model. The demonstration validates the model’s architecture for long-horizon coding agents rather than only single-session inference tasks.
How much does Qwen3.8-Max cost?
Qwen3.8-Max is available through QwenCloud at $2 per million input tokens, $6 per million output tokens, and $0.25 per million implicitly cached tokens. The $0.25 cached token rate is significant for agentic workflows that reuse large context windows repeatedly across long task chains, as it reduces the effective cost of long-horizon runs substantially compared to paying full input rates on every call. No free tier is available at launch. Open-weight release will allow self-hosted deployment at zero per-token cost for teams with sufficient compute, with weights expected to release shortly after August 3, 2026.
Will Qwen3.8-Max be open source?
Yes. Alibaba has committed to open-sourcing the Qwen3.8-Max weights, making it the first Max-class model from any major AI lab to be released with open weights in 2026. A smaller dense model, Qwen3.8-27B, is also planned for open-weight release simultaneously. The weights were not available at the August 3 API launch; Alibaba indicated they would ship approximately one week after the API release. Once available, the weights can be deployed on self-hosted infrastructure, eliminating per-token API costs for teams with sufficient compute.
What is Qwen3.8-Max best used for?
Qwen3.8-Max is best used for long-horizon autonomous coding agents that run multi-day software engineering tasks without human supervision, autonomous machine learning research replication and literature review, enterprise video processing workflows requiring 100-plus-hour video analysis with memory graph construction, and agentic tasks requiring sustained computer environment interaction such as OS-level automation and multi-application workflows. It is less suited for discrete issue-fixing pipelines where Claude Fable 5 holds a meaningful SWE-bench Pro advantage, or for expert-level single-question reasoning where HLE scores favor Fable 5 and GPT-5.6 Sol.
What makes Qwen3.8-Max different from other frontier models?
Three things distinguish Qwen3.8-Max from other frontier models available in August 2026. First, it leads on OSWorld-Verified and PaperBench, the benchmarks most directly measuring autonomous agent task completion rather than single-turn inference quality. Second: it is the only Max-class frontier model with a committed open-weight release, which will enable self-hosted deployment that Claude, GPT, and Grok do not offer. Third, its native video processing capability handles 100-plus-hour streams with continuous memory graph construction, a feature not available at the API level from Anthropic or OpenAI’s current models.
