# No Hype: open weights at a quarter of the price, Lumo, Gemini opt-out

> Open-weight models cost about four times less than closed ones on OpenRouter, Proton ships Lumo for desktop, and Gemini separates chat history from training.

- Published: 2026-11-18
- Author: Jacobo González Jaspe
- Language: en
- Tags: news, opinion, no-hype, ai
- Canonical URL: https://vorluxai.com/en/blog/sin-humo-2026-11-18-en/
- Source: VORLUX AI (J4SGON S.L.), https://vorluxai.com/

---

Open-weight models served 71.7% of the tokens on OpenRouter at the end of September, according to Hugging Face. You pay about four times less for them than for closed models. Hugging Face, OpenRouter, Proton, Google and the US Commerce Department fill this issue.

## Open-weight models are the cheap default

![Illustration generated with AI on our own machine.](https://vorluxai.com/assets/blog/sin-humo/2026-11-18-1.webp)
*Illustration generated with AI on our own machine.*


**What happened.** Hugging Face and OpenRouter are teaming up to make open-weight models easy to run ([Hugging Face](https://huggingface.co/blog/openrouter)). OpenRouter's data shows open-weight models went from 26.8% of tokens served in October 2025 to 71.7% at the end of September 2026. Its measurements show the new routing cut capacity errors by 46% and improved throughput by 16.3%. Open-weight models cost 0.28 dollars per million input tokens on average, against 1.27 dollars for closed models, according to the same post.

**What we think.** We read this as a price list: at 0.28 dollars per million input tokens against 1.27, you ask what your closed model does that earns the difference ([Hugging Face](https://huggingface.co/blog/openrouter)). A router between you and the model is still someone else's server.

**The good bit.** Capacity errors fell by 46% with the new routing, by OpenRouter's own measurements ([Hugging Face](https://huggingface.co/blog/openrouter)).

**What you can do.** Put your monthly token count next to both prices. If the model you rent is open-weight, check with the [VRAM calculator](https://vorluxai.com/en/tools/vram-explorer/) whether it fits on a machine you own.

## Proton brings Lumo to the desktop

![Illustration generated with AI on our own machine.](https://vorluxai.com/assets/blog/sin-humo/2026-11-18-2.webp)
*Illustration generated with AI on our own machine.*


**What happened.** Proton launched Lumo Desktop for Windows, macOS and Linux on Tuesday ([TechCrunch](https://techcrunch.com/2026/11/17/protons-lumo-a-privacy-focused-alternative-to-chatgpt-and-claude-arrives-on-desktop/)). The app includes connectors for Proton's email app and third-party services including Confluence, Dropbox and Jira. Proton says the limited data it stores is protected under Swiss privacy law.

**What we think.** We keep client files out of any assistant that stores data, limited or not; you check what Lumo keeps before you plug in Dropbox or Jira ([TechCrunch](https://techcrunch.com/2026/11/17/protons-lumo-a-privacy-focused-alternative-to-chatgpt-and-claude-arrives-on-desktop/)). Swiss law protects stored data. It does not un-store it.

**The good bit.** One desktop app on Windows, macOS and Linux, with connectors to tools a team already uses.

**What you can do.** Read what Lumo stores before you connect a client folder. For the options that keep everything in your office, see [Private ChatGPT for your business: three options and the cost](https://vorluxai.com/en/blog/private-chatgpt-alternative-for-your-business-en/).

## Gemini separates history from training

![Illustration generated with AI on our own machine.](https://vorluxai.com/assets/blog/sin-humo/2026-11-18-3.webp)
*Illustration generated with AI on our own machine.*


**What happened.** Google is adding more controls to the Gemini app to limit training on user data ([Ars Technica](https://arstechnica.com/ai/2026/11/google-will-let-you-opt-out-of-ai-training-in-gemini-without-losing-your-history/)). The change separates chat history from training, with separate toggles for each. AI training will be an opt-out situation. Google says it may retain data for up to three years to improve its models.

**What we think.** We call an opt-out what it is, a default that trains on your chats until you switch it off; you switch it off, and you remember the three years ([Ars Technica](https://arstechnica.com/ai/2026/11/google-will-let-you-opt-out-of-ai-training-in-gemini-without-losing-your-history/)).

**The good bit.** Chat history and training get separate toggles.

**What you can do.** Find the training toggle when it reaches your account and turn it off. If your team uses Gemini for work, read the [EU AI Act guide](https://vorluxai.com/en/eu-ai-act/) first.

## The US closes the chip rental loophole

![Illustration generated with AI on our own machine.](https://vorluxai.com/assets/blog/sin-humo/2026-11-18-4.webp)
*Illustration generated with AI on our own machine.*


**What happened.** The US Commerce Department said it is closing a loophole that let rivals like China, Russia and Iran rent access to advanced chips from data centres in Indonesia, Malaysia, Singapore, Thailand and Brazil ([Ars Technica](https://arstechnica.com/tech-policy/2026/11/us-moves-to-cut-chinas-remote-access-to-ai-chips/)). The South China Morning Post reported that any advanced chips made for a Chinese firm are assumed to be restricted, and factories must ask the US government for a licence before shipping them. The rule change takes effect 90 days after it is posted to the Federal Register.

**What we think.** We read the rule for what it says about rented compute: access in someone else's data centre is access a government can cut with 90 days of notice; you own the hardware or you depend on the list ([Ars Technica](https://arstechnica.com/tech-policy/2026/11/us-moves-to-cut-chinas-remote-access-to-ai-chips/)).

**The good bit.** The rule comes with a date, 90 days after publication.

**What you can do.** If your plan depends on rented GPUs abroad, write down which country they sit in. For what runs in your own office, see [hardware by job](https://vorluxai.com/en/products/).

## In one line
OpenRouter and Hugging Face put open weights at 0.28 dollars per million input tokens; we compare that price with a machine we own before we rent anything.

## Work with us

If something in this column belongs in your business, on models that run on your own hardware rather than someone else's cloud, [tell us what you need](https://vorluxai.com/en/contact/) or see how we work in [consulting](https://vorluxai.com/en/consulting/).
