Qwen3.8-Max vs Claude Opus 5.5: The Real Numbers

Q: So which should you use?

Qwen3.8-Max for high-volume agentic and terminal work, desktop automation, instruction-heavy pipelines, or anywhere self-hosting is a hard requirement. Opus 5.5 for deep engineering, long refactors, and work where a wrong answer costs more than the tokens. GPT-6 Sol when cost per task is the deciding constraint. Sonnet 5.5 is the one the viral post ignored and arguably the most direct competitor on price. The honest summary: the price gap is real and narrowing the quality gap is real too, but a 3.3x price difference is not the same as equivalence, and a year-old model quoted at the wrong price is not evidence of either. See Alibaba's own benchmark table and the open weights on Hugging Face.

Comparisons

A viral post comparing Qwen3-Max to Claude got the model, the price and the date wrong. Here is the accurate comparison of Alibaba's current flagship against Claude Opus 5.5 and GPT-6 Sol on price, benchmarks and open weights.

AP

Staff

September 29, 2026 · 3 min read

A viral post claiming Alibaba "just launched" Qwen3-Max at 6 dollars per million tokens against Claude at 25 has been circulating this week. Almost every number in it is wrong or a year out of date. But the underlying question is real, because Alibaba’s actual current flagship, Qwen3.8-Max, costs 2 dollars input and 6 dollars output per million tokens against Claude Opus 5.5 at 4 and 20. Here is the accurate comparison.

First, what the viral post got wrong

Qwen3-Max did not just launch. It shipped September 23, 2025, more than a year ago. Alibaba is three generations past it: Qwen3.7-Max in May 2026, Qwen3.8-Max in August 2026.
The 6 dollar figure is not Qwen3-Max’s price. Qwen3-Max runs about 0.78 input and 3.90 output. The 6 dollar output figure belongs to the newer Qwen3.8-Max.
The 25 dollar Claude figure is stale. That was Opus 5. Opus 5.5 launched September 22 at 20 dollars output.
"Rivals Opus 5 on complex reasoning" overstates it. Qwen3-Max scores 19.4 on the Artificial Analysis intelligence index, around the 56th percentile. Not frontier.

What do these models actually cost?

Qwen3.8-Max: 2 dollars input, 6 output. Implicit cache reads 0.25, explicit cache reads 0.17
GPT-6 Sol: 2 input, 10 output, cached input 0.20
Claude Opus 5.5: 4 input, 20 output, cache reads 0.20
Claude Sonnet 5.5: 2 input, 10 output, cache reads 0.20

So the real gap between Alibaba’s flagship and Anthropic’s is roughly 3.3x on output, not the 4.17x the viral graphic claimed. And Sonnet 5.5 matches Qwen on input price exactly.

Does Qwen3.8-Max actually match the frontier?

On some things yes, on the hardest things no. Alibaba’s own published table shows the split clearly. On Terminal-Bench 2.1, Qwen3.8-Max scores 86.6 against 84.6 for both Opus 4.8 and Fable 5, and 88.8 for GPT-5.6 Sol. That is a genuine win over two Claude flagships.

On SWE-bench Pro it scores 67.7 against Fable 5’s 80.0, a 12-point gap on core software engineering. On DeepSWE 1.1 it posts 56.6 against Fable 5’s 70.0 and GPT-5.6 Sol’s 73.0. The pattern holds across the table: competitive on terminal-driven agentic tasks, behind on deep engineering.

It also wins outright in places worth knowing about: 82.8 on IFBench instruction following against 72.7 for GPT-5.6 Sol, 73.2 on PLawBench, 60.2 on HealthBench, and 86.1 on OSworld-Verified for desktop agent work. GPQA Diamond is effectively a tie at 92.6.

The benchmark caveat that matters most

Alibaba ran these evaluations itself, on its own infrastructure, and reports the highest score among harnesses. It notes that Qwen3.8-Max performs best on Claude Code, which is the harness used for its headline Terminal-Bench number. Comparison figures for rivals come from different harnesses and different runs. This is normal practice and it is also why vendor tables consistently flatter the vendor, a problem our guide to model rankings covers.

Note also that the comparison set is Opus 4.8 and Fable 5, not Opus 5.5. On the newer Terminal-Bench 4.0, Opus 5.5 leads at 66.4% while Qwen3.8-Max sits around 67.7 on a differently-run board. Cross-version comparisons here are not clean.

The thing the viral post missed entirely

Alibaba released the weights. Qwen3.8-2.4T-A95B went public on August 12, 2026: 2.4 trillion parameters with 95 billion active per token, 262K native context extensible to about 1.01 million. The open build omits image input and non-thinking mode, but it is a genuinely frontier-adjacent model you can self-host.

That matters more than a price comparison, because it removes the vendor from the equation entirely for regulated environments that cannot call third-party APIs. It is the same dynamic behind AT&T routing 40% of its AI traffic to open models and DeepSeek shipping V4.1 Flash under MIT.

So which should you use?

Qwen3.8-Max for high-volume agentic and terminal work, desktop automation, instruction-heavy pipelines, or anywhere self-hosting is a hard requirement. Opus 5.5 for deep engineering, long refactors, and work where a wrong answer costs more than the tokens. GPT-6 Sol when cost per task is the deciding constraint. Sonnet 5.5 is the one the viral post ignored and arguably the most direct competitor on price.

The honest summary: the price gap is real and narrowing the quality gap is real too, but a 3.3x price difference is not the same as equivalence, and a year-old model quoted at the wrong price is not evidence of either. See Alibaba’s own benchmark table and the open weights on Hugging Face.

Claude Sonnet 5.5 Beats Opus 5.5 on Terminal-Bench for Half the Price

Claude

Anthropic released Claude Sonnet 5.5 at unchanged Sonnet 5 pricing of $2/$10, scoring 70.6% on Terminal-Bench 4.0 against Opus 5.5's 66.4%, and adding cyber limits previously reserved for frontier models.

AP

Staff

September 29, 2026 · 2 min read

Anthropic released Claude Sonnet 5.5 on September 28, 2026, six days after Opus 5.5 and the second model in the 5.5 family. The price did not move: 2 dollars per million input tokens and 10 per million output, the same as Sonnet 5. Anthropic says it runs more than 30% faster and costs up to 30% less per task. The more interesting number is that it scores 70.6% on Terminal-Bench 4.0 against Opus 5.5’s 66.4%, while costing half as much per token.

Wait, the cheaper model beats the expensive one?

On that specific benchmark, at that specific setting, yes. Sonnet 5.5’s 70.6% on Terminal-Bench 4.0 is above Opus 5.5’s 66.4% at its highest effort setting. Both headline scores use the most expensive configuration, which is the caveat that matters. Terminal-Bench measures agentic command-line work, not general reasoning, so this is not a claim that Sonnet is smarter overall.

Still, it is a real result and it mirrors what happened at DeepSeek, where V4.1 Flash outperformed the larger V4-Pro flagship on coding. Cheaper models beating expensive ones on agentic tasks is becoming a pattern rather than an anomaly.

Where does the 30% saving come from?

Not from the price, which is unchanged. The claim is that Sonnet 5.5 needs fewer tokens to finish the same work, so the per-task bill drops even at the same per-token rate. That is a real effect if it holds on your workload, and it is worth measuring rather than assuming. Artificial Analysis measured higher task costs than Sonnet 5 at maximum effort, which cuts against the headline.

The full rate card: 2 dollars input, 10 output, 0.20 cache reads, 2.50 cache write at five minutes, 4 dollars cache write at one hour, and 50% off on the Batch API. Opus 5.5 sits at 4 and 20. Context window is 1 million tokens with 128,000 max output. Available on the Claude API, Amazon Bedrock, Google Cloud and Microsoft Foundry as claude-sonnet-5-5.

One pricing detail worth knowing

Anthropic describes 2 and 10 as the same pricing as Sonnet 5, and that is accurate, but there is history. Those rates were originally sold as introductory pricing due to rise to 3 and 15. Anthropic made them permanent instead. So the comparison is honest, though it is a comparison to a promotional rate that stuck rather than to an original list price.

The safety addition nobody is talking about

Sonnet 5.5 is the first Sonnet to ship with the cyber restrictions Anthropic previously reserved for its most capable models, and the first with classifiers that stop its reasoning being extracted. That second part has a specific trigger: researchers recently decoded 315,320 thinking blocks from publicly posted agent traces. Reasoning traces leaking is a real exposure, and this is a direct response.

Applying frontier-tier cyber limits to a mid-tier model also fits the restricted-access direction seen in Google’s Fairwind Program. One awkward footnote: Claude suffered a partial outage the day after this launched, which is the kind of thing that makes a two-models-in-seven-days cadence worth watching.

See Anthropic’s announcement.

Loading more stories…

You've reached the end — that's every story we've published.