Qwen 3.6 Plus Beats DeepSeek V4 Pro in Russian Battle Test and Costs Less

A fresh battle test of six April LLMs on Russian content produced an unexpected result: DeepSeek's new flagship, V4 Pro, did not come out on top. Qwen 3.6 Plus beat it, even though it was released earlier and costs less.

Who came out ahead

DeepSeek V4 Pro was expected to reach Tier S, meaning above 95 points out of 100. The expectation made sense: the model is large and new, posts strong results on AIME and SWE-bench, and is built around a reasoning architecture. But in the practical test on Russian content, it scored 89 points. That is a high score, but not the one usually expected from a release that claims flagship status, particularly for this market.

The comparison within the DeepSeek lineup is even more interesting. The Flash version scored 83 points, only 6 points behind Pro. Against that backdrop, a repeat run of Qwen 3.6 Plus, which was released 22 days earlier, produced 92 points. The older model thus beat DeepSeek's newest release not only on text quality but also on overall usefulness for real tasks in Russian.

Price versus quality

The main surprise is not just the scores but the economics. If Pro beats Flash by only a few points while costing 13 times more, the production decision is no longer obvious. For teams that generate large volumes of content, that difference quickly turns into a significant line item. In that scenario, what matters is not the absolute record in a test but how much useful output the model delivers for every dollar spent.

In the updated methodology, the author of the comparison focuses on exactly this and proposes looking at models through score-per-dollar. This approach changes the conclusions more than a standard ranking by raw scores does. A model may be slightly behind on quality but win in real-world use thanks to its price, speed, and more predictable behavior on long responses. For editorial teams and product teams, this is far more useful than blindly paying for the most expensive option.

• DeepSeek V4 Pro: 89 points against an expected Tier S
• DeepSeek Flash: 83 points and noticeably more forgiving economics
• Qwen 3.6 Plus: 92 points and the lead in the comparison
• Difference between Pro and Flash: 6 points against a 13x price difference
• Key metric for choosing a model: not only score, but also score-per-dollar

Why reasoning didn't save it

One of the main hypotheses after the test is that optimizing for reasoning does not guarantee a strong narrative result. Metrics such as AIME and SWE-bench show a model's ability in math, code, and structured reasoning, but they are weaker predictors of how well it writes lively, coherent, and persuasive text in Russian. Content tasks depend on rhythm, precise wording, a sense of structure, and handling the nuances of the language, not just the ability to break a problem down into the right steps.

Against this backdrop, the methodology updates look less like a formality and more like an attempt to bring the test closer to production in an honest way. The changes include a max_tokens setting, a paid re-test, and a stricter assessment of the practical value of an answer. In other words, the comparison no longer covers merely "smart" models, but models that must reliably handle a specific editorial task within a set budget. Under exactly these conditions, it turned out that a release being new is no longer an advantage in itself.

What this means

The LLM market looks less and less like a race where whoever is newest is best. For Russian-language content tasks, the winner is not the loudest model but the one that holds text quality better and pays off in production. For teams, this is a signal to re-check fresh flagships on their own scenarios more often, rather than choosing them based only on benchmark headlines.