This Week in AI, 15 Items — Big Open-Weight Models Arrive Together, and Europe Moves on AI for Science

If you try it through the API, budget for $2 / $10 as an introductory price that doubles on 1 December. It is not available on the free tier. The safety numbers are Google’s own evaluations, and the cyber-capability access for defenders (the Fairwind Program) continues as a separate track.

This is the issue where Meta joins the list of models you can call over an API. But with no price or numbers in the official text, I would hold off on comparing or switching until you have read the price list and the benchmark tables directly. The performance claims are Meta’s own.

Building it into your own product or service is broadly allowed on paper if you keep the attribution and follow the Usage Policy. The condition to watch is offering it to others through an API (a separate agreement above US$20 million in revenue). The weights are gated, about 1.5 TB in BF16, so few environments can run them. For a legal decision, read the original terms rather than this summary.

Apache 2.0 gives you a licence to set against the custom terms of the previous item (Meta). It is text-only, the performance claims are Reflection’s own, and running it yourself takes 501B-class resources.

There are now 24 GB and 4 GB options for agents you run locally. The speed and ranking claims are NVIDIA’s own comparisons. I did not check the licence terms this time.

It gives you a window to try Japanese-built models through an API. It is free and quick with a Google account. Because of the note about user data being used for quality improvement, it is safer to keep confidential or personal data out of any trial (my judgement from that note). With no performance figures published, you have to test it against other models yourself.

If you work with time series such as demand forecasts, sensor monitoring or capacity planning, this is a new baseline you can try without a GPU. The rankings are Mistral’s own submitted results, which I did not check against the leaderboards, and I did not check what “Modified MIT” changes.

A cloud company, PFN and a public research institute are dividing the work of putting continuous checks of Japanese models against Japanese law and culture into practice. But the start of operation is a target for next fiscal year, and this is still a validation stage. I could not find in the release how evaluation criteria or results will be published.

An example is appearing of using a domestic LLM, without an external cloud, where sensitive information is handled. Results will not be out until after March 2027, so there are no effect figures yet. The announcement does not describe tsuzumi 2’s performance itself.

This is a window through which European research bodies can receive private AI and compute resources. It targets researchers, not general developers. Google’s $3 million is the only amount stated, so the real scale will not be clear until the progress reports.

If you leave long conversations idle, you avoid the cache-expiry reprocessing cost, but remember the conversation is compacted. If you use claude.ai/design, you need to decide where to move before 14 December (the in-tool notice is the only basis; I did not check an Anthropic announcement).

The question is whether METR’s publication lets outsiders see how much was shown. Anthropic judges which information falls under the redaction categories (per its own document). METR’s analysis is not out yet, so the substance of the evaluation is unconfirmed.

The material here is a chatbot maker publishing numbers on how people use it around elections. But it is one week of conversations classified automatically by Claude, and it only shows that requests for voting recommendations are very rare. I did not check whether other companies publish comparable figures.

It backs the view that agentic use is becoming the norm, using OpenAI’s own usage data. But it covers only OpenAI’s business customers, not the EU’s companies as a whole, and “frontier” is its own definition (top 10% by token volume).

It is a statement that Microsoft’s own products will lean on its own chips for inference. All figures are Microsoft’s design targets or self-comparisons, and Maia 300 has not shipped. The announcement does not give the price or terms for customers running on Maia.