Open vs closed models: the game that matters more than the leaderboard

Every time a new model comes out, the question is the same: is it top of the leaderboard? I think for most people and companies it’s the wrong question. The right one is different: can I control it?

Where things stand

In October 2026 the most capable models are closed: you use them through apps or APIs but can’t download them. On the Artificial Analysis index the leader, Claude Opus 5.5, scores 58. The best open-weights model, Xiaomi’s MiMo-V2.6-Pro, scores 46; GLM-5.3 and Kimi K3 follow. There is a real gap.

But two facts tell another story. MiMo-V2.6-Pro’s cost per task on that benchmark is 13 cents, against almost $6 for the leader. And today’s open scores are what closed models scored not long ago: the gap is measured in months, not years.

Why open models matter

Data control. An open model runs on your servers. For a professional firm, a hospital or a company with trade secrets, that’s the difference between being able to use AI on certain documents or not.

Independence. If a closed model’s provider changes prices, terms or retires a model, you change with it. An open model downloaded today will work the same in five years.

Specialisation. An open model can be adapted to your vocabulary, documents and language. For many tasks a specialised mid-size model beats a huge generic one.

Competition. Open models keep everyone’s prices down. Every time a lab releases competitive weights, closed ones must justify their price.

Why closed models stay ahead

Building the best models costs billions in compute, data and researchers. Closed labs can monetise directly and integrate the model into polished products: apps, agents, tools. And for the hardest tasks (long agentic work, complex code, advanced reasoning) the difference is still felt.

There’s also safety: an open model can’t be “recalled” if a dangerous use is discovered. It’s a serious argument, which open labs address with pre-release testing and usage licences.

My position: both, with judgement

For a professional or SME the answer isn’t ideological:

Closed for hard, non-confidential tasks where top quality makes the difference: complex analysis, important writing, agents.
Open for repetitive tasks on confidential data, where control and cost matter: classifying documents, extracting data, internal assistants.
Portable processes: prompts, instructions and data organised so you can switch models without starting over.

The mistake to avoid is building everything around one provider as if it were eternal. Technology history teaches that today’s winners aren’t guaranteed tomorrow.

In one line

The leaderboard says who’s best today. Open vs closed decides who controls your data tomorrow.

Sources

Informational content, current as of the date shown. Prices, models and rules change often: always check official sources. Written with the support of AI tools and reviewed by the author.