For two weeks, the Qwen 3.8 story had a gap: July press previews promised a frontier-class model but shipped no scores. On August 3, 2026, Alibaba published its official Qwen 3.8-Max release post, including benchmark tables comparing it with Claude Opus 4.8, Fable 5, GPT-5.6 Sol, Qwen3.7-Max, Gemini 3.1 Pro, and GPT-5.6 Sol.
The tables show real strengths, clear losses, and important methodology caveats. This post breaks down the published results, then shows how to evaluate Qwen 3.8-Max against your own API workload instead of relying on vendor benchmarks. For model parameters, pricing, and access details, see the Qwen 3.8-Max explainer.
All scores below come from Alibaba’s published table, with leaderboard footnotes dated August 3, 2026. No independent evaluations were available at the time of writing, so treat each number as a vendor-reported result.
The table at a glance
Alibaba’s headline text-model results are below. Higher is better for every benchmark.
| Benchmark | Qwen 3.8-Max | Claude Opus 4.8 | Fable 5 | GPT-5.6 Sol | Qwen3.7-Max |
|---|---|---|---|---|---|
| Terminal Bench 2.1 | 86.6 | 84.6 | 84.6 | 88.8 | 74.5 |
| SWE-bench Pro | 67.7 | 69.2 | 80.0 | 64.6 | 60.6 |
| PaperBench | 93.0 | 80.3 | 88.8 | 90.5 | 64.8 |
| GPQA Diamond | 92.6 | 92.0 | 92.6 | 94.1 | 92.4 |
| IFBench | 82.8 | 62.2 | 63.5 | 72.7 | 79.1 |
| HLE (Humanity’s Last Exam) | 43.6 | 45.7 | 53.3 | 47.2 | 41.4 |
Source: Alibaba’s vendor-run benchmark table.
Where Qwen 3.8-Max wins
PaperBench: strongest flagship result
On PaperBench, which measures whether a model can reproduce research-paper results, Qwen 3.8-Max scores 93.0:
- GPT-5.6 Sol: 90.5
- Fable 5: 88.8
- Claude Opus 4.8: 80.3
- Qwen3.7-Max: 64.8
The jump from Qwen3.7-Max is 28.2 points. That is substantial and should be validated independently, but it suggests a meaningful improvement in long-horizon research tasks if the result holds.
IFBench: instruction following
Qwen 3.8-Max scores 82.8 on IFBench. The nearest non-Qwen model in Alibaba’s table is GPT-5.6 Sol at 72.7.
| Model | IFBench |
|---|---|
| Qwen 3.8-Max | 82.8 |
| Qwen3.7-Max | 79.1 |
| GPT-5.6 Sol | 72.7 |
| Fable 5 | 63.5 |
| Claude Opus 4.8 | 62.2 |
Since Qwen3.7-Max already led this row, instruction following appears to be a sustained Qwen strength rather than a one-generation result.
Terminal Bench 2.1: ahead of Claude, behind GPT-5.6 Sol
Alibaba reports 86.6 for Qwen 3.8-Max on Terminal Bench 2.1:
- GPT-5.6 Sol: 88.8
- Qwen 3.8-Max: 86.6
- Claude Opus 4.8: 84.6
- Fable 5: 84.6
This is a second-place result, not an overall win. Still, it puts Qwen 3.8-Max ahead of both Claude Opus 4.8 and Fable 5 in Alibaba’s agentic terminal evaluation.
Multimodal performance
The separate multimodal table is Qwen 3.8-Max’s strongest overall showing. Alibaba reports:
- MathVision: 95.2
- LogicVista: 91.9
- OSWorld-Verified: 86.1
The model also leads nearly every OCR row in the published table. Claude Opus 4.8 and Fable 5 are text-focused in this comparison, so they do not directly compete across these multimodal benchmarks. Against Gemini 3.1 Pro and GPT-5.6 Sol, visual reasoning and document intelligence are Qwen 3.8-Max’s clearest reported differentiators.
For the comparison with the other major open-weight summer release, note that Kimi K3 is text-only. See the full Qwen 3.8 vs. Kimi K3 comparison.
Where it loses
Alibaba included weaker rows in the table. Three are especially relevant for developers evaluating coding and general-reasoning workloads.
HLE: behind every flagship competitor
On Humanity’s Last Exam (HLE), Qwen 3.8-Max scores 43.6:
| Model | HLE |
|---|---|
| Fable 5 | 53.3 |
| GPT-5.6 Sol | 47.2 |
| Claude Opus 4.8 | 45.7 |
| Qwen 3.8-Max | 43.6 |
| Qwen3.7-Max | 41.4 |
Qwen 3.8-Max improves on its predecessor, but it ranks last among the four flagship models in this table. Since HLE is a broad-knowledge benchmark designed to resist narrow benchmark tuning, this result deserves significant weight.
SWE-bench Pro: far behind Fable 5
On SWE-bench Pro, Qwen 3.8-Max scores 67.7:
- Fable 5: 80.0
- Claude Opus 4.8: 69.2
- Qwen 3.8-Max: 67.7
- GPT-5.6 Sol: 64.6
- Qwen3.7-Max: 60.6
The model improves meaningfully over Qwen3.7-Max, but it is 12.3 points behind Fable 5. Both statements can be true at once:
- Qwen 3.8-Max beats Claude Opus 4.8 on Terminal Bench 2.1.
- Qwen 3.8-Max trails Fable 5 substantially on SWE-bench Pro.
DeepSWE and difficult agentic coding tasks
Alibaba’s own table also shows Qwen 3.8-Max behind frontier models on DeepSWE. The coding pattern is consistent:
- Competitive on terminal-driven agent tasks.
- Less competitive on the deepest software-engineering evaluations.
Read the methodology before using the numbers
Vendor benchmark tables are useful, but they are not independent measurements. Four details from Alibaba’s publication affect how these results should be interpreted.
1. Every result is vendor-run
Alibaba evaluated Qwen 3.8-Max and its competitors. This is common for launch announcements, but vendors control details that can affect results:
- Benchmark version
- System prompt and prompting strategy
- Sampling settings
- Retry policy
- Tool configuration
- Agent harness
That does not make the results invalid. It means they should be treated as claims that require confirmation.
2. Most coding benchmarks used the Claude Code harness
Alibaba ran most coding benchmarks with Claude Code, Anthropic’s agent harness, connected to Qwen 3.8-Max through an Anthropic-compatible API.
This is useful because the models were evaluated in a shared tool environment. However, agentic benchmarks are harness-sensitive. A score in this setup should be read as:
Qwen 3.8-Max running inside the Claude Code harness
—not as a harness-independent coding capability score.
3. Several benchmarks are Qwen-built
QwenSWEBench, QwenQoderBench, CoWorkBench, and RecreationBench were built by the Qwen team.
In-house benchmarks can capture capabilities that public benchmarks miss. However, they are different evidence from public evaluations such as SWE-bench Pro. When using a score from the table, identify whether the benchmark is public or vendor-built first.
4. The Fable 5 footnote matters
Alibaba’s table notes that:
“Fable5 results may involve fallbacks.”
That means the Fable 5 column may not represent clean, direct model execution in every case. Since Fable 5 is often the model Qwen 3.8-Max trails in the table, this caveat matters when comparing exact deltas.
This methodology issue is not unique to Alibaba. Anthropic, OpenAI, and Google also publish self-run benchmark tables. The same pattern appeared in the Kimi K3 benchmarks breakdown: vendor tables are useful starting points, not final measurements.
How to evaluate the next vendor table
Use this checklist before adopting benchmark claims:
Who ran the evaluation?
Vendor-reported competitor numbers should receive extra scrutiny.Which harness and settings were used?
Tool use, retry behavior, prompts, and agent harnesses can change coding results.Which benchmarks are vendor-built?
In-house evaluations may be useful, but they are not equivalent to public benchmarks.Did you read the footnotes?
Notes such as “may involve fallbacks” can materially change how a result should be interpreted.What is missing?
Compare current tables with earlier generations to find rows that disappeared. The previous cross-vendor comparison can help identify missing benchmarks.Is there an independent second source?
Waiting for community leaderboards or third-party evaluations often provides more actionable evidence than a launch-day table.
Run your own evaluation instead
A public benchmark can help shortlist models. Your production prompts should make the final decision.
Qwen 3.8-Max is available through Alibaba Cloud Model Studio as qwen3.8-max, listed on the official models page. It supports both OpenAI-compatible and Anthropic-compatible APIs.
A practical evaluation workflow in Apidog looks like this:
- Create a request for Qwen 3.8-Max using the Model Studio chat-completions endpoint.
- Create an identical request for your baseline model, such as Qwen3.7-Max, Kimi K3, Claude, or GPT.
- Save 20–30 representative production prompts as a test scenario.
- Add assertions for the properties your application requires:
- Valid JSON
- Required fields present
- Correct schema
- No unsupported tool calls
- Latency below your threshold
- Run both scenarios with the same inputs.
- Compare pass rate, output quality, response time, and cost.
For Qwen 3.8-Max specifically, test two details:
- The model defaults to
reasoning_effort: xhigh, so benchmark the reasoning level you would actually deploy. - If you plan to use the Anthropic-compatible endpoint, test it separately. That protocol shape is also the one Alibaba used for most of its coding benchmarks.
Use Apidog to configure each endpoint as an environment and generate side-by-side test reports for your actual workload. You can download Apidog and start collecting first-party data before independent leaderboards publish theirs.
Frequently asked questions
Are the Qwen 3.8 benchmark numbers independently verified?
No. As of August 3, 2026, the published numbers come from Alibaba’s own table. Artificial Analysis and community leaderboards had not yet published Qwen 3.8-Max scores at the time of writing. Treat the results as vendor-reported until independent evaluations are available.
What is Qwen 3.8-Max’s best benchmark result?
Its strongest flagship-table row is PaperBench at 93.0, ahead of GPT-5.6 Sol at 90.5, Fable 5 at 88.8, and Claude Opus 4.8 at 80.3. In the multimodal table, MathVision at 95.2 and the OCR results are its strongest reported results.
Where does Qwen 3.8-Max clearly lose?
Qwen 3.8-Max scores 43.6 on HLE, ranking last among the four flagships and behind Fable 5’s 53.3. It also scores 67.7 on SWE-bench Pro versus Fable 5’s 80.0, and trails on DeepSWE. Deep software-engineering evaluations and broad-knowledge exams are its weaker areas in Alibaba’s table.
How much better is Qwen 3.8 than Qwen 3.7-Max on benchmarks?
Alibaba reports substantial generation-over-generation gains:
- Terminal Bench 2.1: 74.5 → 86.6
- SWE-bench Pro: 60.6 → 67.7
- PaperBench: 64.8 → 93.0
Whether the upgrade is worthwhile depends on your workload, modality requirements, latency targets, and price. See the full Qwen 3.8 vs. Qwen 3.7 comparison.
Top comments (0)