In short
- Mistral Medium 3.5 is a 128 billion parameter thick model at a price of $1.50 per share / $7.50 output per million tokens, far from comparing China’s method.
- Chinese open source models – Qwen, GLM, MiMo-V2 – dominate the top, leaving Mistral as the only Western model.
- Mistral is positioning the release as a building block for a larger version of the future.
Mistral AI dropped Mistral Medium 3.5 on April 29. The Paris-based lab announced a 128-billion-parameter model, a set of agents – and it went straight to the wall of “meh” on the Internet.
The release came in three phases. First, the example itself. Second, remote coding agents via the Mistral Vibe CLI—cloud-based components that can push requests to GitHub and run them together without you being at the terminal. Third, the Work Method CatMistral’s ChatGPT consumer interface, which now operates autonomously across multiple steps such as email analytics, survey synthesis, and tool navigation.
Big ambitions, but a disturbing reality.
3.5 average scores 77.6% on SWE-Bench Verified—a test benchmark that tests whether a version can fix real GitHub issues by creating functional patches. It also hits 91.4% on τ³-Telecom, which tests the use of military equipment in special areas. Mistral has combined three different models in the past (Medium 3.1, Magistral, and Devstral 2) into one set of weights and variable effort on request.
A hybrid model instead of three is a real technological breakthrough. The problem is what is the price.
Mistral charges $1.50 per million tokens input and $7.50 per million tokens output. Alibaba’s Qwen 3.6 at 27 billion units – less than a quarter of Medium 3.5’s parameter count – scored 72.4% on the verified SWE-Bench benchmark and ships under Apache 2.0, meaning you can download and run it for free.
Did you know?
Parameters are what determine the AI’s ability to learn, reason, and store information. The more parameters, the greater the knowledge of the model.
Take a look at the open source leaderboards and the picture is perfect. The top spots are Alibaba’s Qwen, GLM from China Zhipu AI, and MiMo-V2 from Xiaomi, all of which are cheaper, more powerful and more competitive than the new Mistral release. The average 3.5 has never been on the major independent billboards—third-party reviews still exist.
The only good thing, as others say, is that Mistral is, at the moment, the only non-Chinese who has any problem with open negotiations.
The Internet does something
Pedro Domingos, a professor of machine learning at the University of Washington, was not complacent:
“Mainstream AI companies boast that their model is the best in benchmarks. Only Mistral boasts that it’s the worst.”
He followed up with a sharp question: “I don’t know what is worse, for Europe not to be in the AI race or to stand and laugh like Mistral.”
Youssof Altoukhi, founder of Yoyo Studios, he did the math: Qwen 3.6, at 27 billion units, is 4.7 times smaller than Medium 3.5 and much the same for coding. An average output rate of 3.5 puts it alongside closed-loop models as the highest of any major benchmark.
“If it wasn’t for their political skills, they would have no money by now,” he said.
Not everyone was proud. AI expert Michal Langmajer quotes:
“I’m really glad there’s a non-US, non-Chinese lab trying to make borderline LLMs but guys we need to step up the game in Europe. Their new model is ‘unsuccessful’ by any benchmark, yet they cost several times more than most of their competitors.”
Some developers have pointed out that the open scale is a solid game, not a board game. A model that anyone can download, improve, and create for themselves doesn’t need success these days to be relevant. Some have pointed to the deployment of real Mistral businesses across Europe as proof that moat is more than just technology.
Geopolitical safety net
This is where the Mistral actually resides.
European businesses subject to GDPR, banks that deal with customer service, and governments that cannot run AI services through Chinese infrastructure have limited options. Like Decrypt report Last December, HSBC signed a multi-year contract with Mistral specifically to build its own models for its infrastructure. The appeal of the open heavy lab in the capital of the EU with a cost of $ 14 billion is not visible in the standards – but it is visible in the purchase decisions.
Not the best of writing, and cheap. But here it is: not American, not Chinese, limited, self-contained, and legally safe from European companies.
Daily explanation A letter
Start each day with top stories right here, including originals, podcasts, videos and more.
