Qwen3.8-Max: €4.52 a Session vs €26 for Claude Fable 5 — Where Cheap Gets Expensive

What launched, in one paragraph

The official announcement went live at 10:00 Beijing time on 3 August 2026. Qwen3.8-Max is a Mixture-of-Experts model: 2.4 trillion total parameters, 95 billion active per forward pass. It has a 1M-token context window, accepts up to 991K input tokens and emits up to 131K. It is available now through QwenCloud, and Alibaba says the weights will be published next week — the first time a Max-class Qwen model has gone open-weight. We flagged this release as one to watch in our August 2026 model preview; now the numbers are on the table.

The price list, in euros

Vendors publish in dollars, European finance departments budget in euros. Using the Czech National Bank reference rates for 3 August 2026 (1 USD = 0.8706 EUR), here is what the four models in Alibaba's own comparison table actually cost:

On output tokens — the expensive half, and the half agentic workloads burn through — Qwen3.8-Max is 4.2× cheaper than Claude Opus 4.8, 5.0× cheaper than GPT-5.6 Sol, and 8.3× cheaper than Claude Fable 5. During the preview period Alibaba is charging 10% of list, which puts input tokens at roughly 17 euro cents per million.

What a real session costs

Per-token prices are abstract. Let me price something concrete: a single agentic coding session that reads a decent chunk of a repository and writes code — call it 2 million input tokens and 200,000 output tokens. That is a normal afternoon for a coding agent, not a stress test.

This is the number I find genuinely useful, because it converts directly into a retry budget. If Qwen3.8-Max needs three attempts to produce what Claude Opus 4.8 produces in one, you have broken even. Against Claude Fable 5 you can afford 5.8 failed attempts before the cheaper model stops being cheaper. That reframing matters: the question is never "is it as good", it is "is the quality gap smaller than the price gap".

So how big is the quality gap?

Here is where it gets interesting, because I went through all 31 rows of Alibaba's headline benchmark table and counted outright wins rather than reading the summary. Qwen3.8-Max leads in six of them — PaperBench, WideSearch, IFBench, HealthBench, PLawBench and PRBench-Finance — and ties for first in a seventh.

In the twelve-row "Coding Agent" section, the section the launch is named after, it wins exactly one: PaperBench, at 93.0. And there is a detail I did not expect. Alibaba included four of its own in-house benchmarks in the comparison — QwenSWEBench, QwenQoderBench, QwenReactBench and QwenSVGBench. Qwen3.8-Max loses all four. Claude Fable 5 takes the first three (86.3 vs 80.7; 63.1 vs 58.4; 1770 vs 1724 Elo) and GPT-5.6 Sol takes the fourth (1758 vs 1713). When a model does not top the benchmarks its own creators designed, "a new bar for coding" is doing heavy lifting.

The worst gaps are worth naming, because they are where the retry budget gets spent: DeepSWE 1.1 (56.6 against GPT-5.6 Sol's 73.0), SWE-bench Pro (67.7 against Fable 5's 80.0), and MLS-Bench-Lite (41.0 against 49.9). On complex software engineering, a 12-to-16-point deficit is not something three retries reliably closes.

The value calculation, and its limits

Take Terminal Bench 2.1, a reasonable proxy for agentic coding, and divide the score by the euro cost of a million output tokens:

Qwen3.8-Max returns roughly 4.3× more benchmark points per euro than Claude Opus 4.8 and 8.6× more than Claude Fable 5, while finishing 2.2 points behind the leader. I should be honest that this metric is crude — benchmark points are not linear, and the difference between 84.6 and 88.8 on a hard agentic test can represent a category of task the cheaper model simply cannot finish. Treat it as a starting point for a procurement conversation, not a verdict.

The part the pricing debate misses

If you only read the coding numbers you will reach the wrong conclusion about where this model is strong. In the multimodal table, Qwen3.8-Max leads nine of twelve visual reasoning rows — MathVision 95.2, LogicVista 91.9, HiPhO 90.0, SLAKE 90.8. On BabyVision it scores 82.0 where Claude Opus 4.8 manages 28.4. In document intelligence and OCR it wins almost everything: OmniDocBench 92.1, CC-OCR-Bench-V2 79.6, MADQA 91.8.

For a European business processing invoices, scanned contracts, or technical drawings, that is a far more relevant result than another point on SWE-bench — and it is available at a fifth of the price of the nearest Western competitor. The model is undersold by its own marketing.

Open weights, EU hosting, and the licence nobody has seen

The weights are the other half of the cost story, because self-hosting changes the economics and the compliance picture at once. Running the model inside the EU removes the cross-border transfer question that a Chinese-hosted API raises under GDPR, and gives you full control over retention.

The catch is scale. I ran the arithmetic: 2.4 trillion parameters at 4-bit quantisation is roughly 1.2 TB, and double that in FP8. That is about fifteen H100-class GPUs with 80 GB each — a data-centre commitment, not a workstation. The companion Qwen3.8-27B is the one individuals and small teams can actually run; at 4-bit it lands around 16–17 GB, which is tight on a 16 GB card and comfortable on 24 GB.

That threshold matters more than it sounds. In our own AI Arena measurements on an RTX 5060 Ti 16 GB, a dense 32B Qwen at 4-bit needs roughly 19–20 GB, overflows the card, and collapses to 4.2 tokens per second — while a similarly sized MoE Qwen that fits runs at 36.3 tokens per second across 157 recorded runs. Overshooting VRAM does not cost you a few percent; it costs you an order of magnitude.

The bigger unknown is legal. Qwen 3.5 and 3.6 shipped under Apache 2.0, which is genuinely permissive. For Qwen3.8-Max no licence file and no model card have been published, and a restrictive community licence would technically satisfy "open weights" while gutting commercial use. Until that file appears on Hugging Face, treat commercial deployment plans as unconfirmed.

There is an AI Act dimension too. General-purpose AI models carry obligations to maintain and supply technical documentation — training methodology, evaluations, and information for downstream providers who build on the model. Chinese vendors have historically been sparing with training-data disclosure, so European deployers should document what they can obtain before the model reaches production, not after.

The verdict

Qwen3.8-Max is not the best coding model available, and Alibaba's own benchmark table says so more clearly than any competitor could. What it is, is the best value per euro by a wide margin in a market where the frontier has been expensive by default — and quietly the strongest publicly available model for vision, documents and OCR. If your workload is document-heavy, the pricing makes the decision easy. If it is complex software engineering, budget for the retries before you switch.