Parameter Update: 2026-25

This week: Loads of Anthropic news, and Meta showcasing that they actually know how to do impressive AI research!

Anthropic

Mythos is back

After weeks of back-and-forth with the US government, Anthropic was allowed to roll Fable 5 back out - to everyone, not just US citizens. Unfortunately, they made the safety classifiers even more aggressive than before, so it's effectively useless for most of my relevant work. Very excited to see how long it takes the Chinese labs to work their ways around the filters and distill it - otherwise things will get very expensive when the model goes back to full API pricing tomorrow.

Sonnet 5

On paper, Sonnet 5 seems like a good launch. It almost matches Opus 4.8 performance at slightly more than half the price.

Unfortunately, these benchmarks are a bit misleading. While performance is good, the model uses a huge amount of tokens to get there. So even at cheaper per-token pricing, the model ends up more expensive, slower, and worse than Opus 4.8. Skip this one.

Claude Science

It's a shame Anthropic refuses to support my specific research use cases, because Claude Science seems like exactly the type of tool I would love to use for them. It obviously falls into the realm of "vibecoded apps built to support a specific persona some GTM person at Anthropic came up with" that also spawned Cowork and Design, but given it's aimed at researchers, I'll not complain.

Direct integration with scientific databases, the ability to track experiments and auto-review work, and other domain-specific features all seem designed to compete with OpenAI Prism (remember OpenAI's LaTeX editor? It seems that they don't either!).

Google

Nano Banana 2 lite

Image generation models have gotten really good over the past year. Two of the biggest remaining criticisms one might levy against them are price and latency. Nano Banana 2 Lite is half the price of Nano Banana 2 (which makes it roughly 1/4th the price of Nano Banana Pro), while also being 4x cheaper.

In my (admittely brief) testing, Nano Banana 2 Lite is still great at most of the things people care about (text correctness, consistency, faces), but a little worse at the things that are harder to quantify (artstyle). In that regard, it follows on the footsteps of Nano Banana 2 that, while winning most benchmarks against Nano Banana Pro, loses out in exactly these qualities.

On the other hand, even if it's a slightly worse tool - the fact that you can try 5x as many ideas as before at the same time might lead to better results?

Gemini Omni Flash

With this one, I was fully ready to give you the same spiel as above - "it's a cheaper version of the big thing, go on, have fun". Then I realized that Google never actually launched an API for the "big" Gemini Omni model?

This means the only real comparison points we have are their own Veo models, and the big Chinese video models. Compared to the former, it's a big performance upgrade at the same price ($0.10 per second). Compared to the latter, it's a good performance jump at a very, very competitive cost.

There's two reasons this release is surprising to me:

1. Given the "Omni" branding, I would have expected more modalities (input, and, more interestingly, output) before the launch of a "Flash" version. Instead, they openly talk about how certain modalities are broken in the API at launch.
2. Given this appears to be the strongest and cheapest video generation model on the market right now, I am surprised at the lack of noise they're making around it. Apart from the launch blog (which led with the new Nano Banana model in the same post?!), they really aren't talking about it as much as I would've expected?

Either way, it's a very fun model that makes it very clear that video generation is getting close for primetime.

Meta Brain2Qwerty2

While there is a lot to be said about Meta's ongoing LLM struggles, morale struggles, leadership struggles, and the like, no one should claim that they aren't capable of producing really solid research.

The same week, they announced Brain2Qwerty v2, follow-up to their original Brain2Qwerty work that has just been published in Nature. They also just open-sourced the original dataset.

The model is trained on MEG and EEG data of nine people typing short sentences on a keyboard. The two innovations from Brain2Qwerty v1:

• Massively more data: 22,000 vs 2,000 sentences, which should lead to better results due to scaling laws if nothing else
• Continuous vs. character-level decoding: Instead of just decoding individual characters given highly aligned labels, the pipeline is designed to work on contiuous data end-to-end (i.e., brain signal in, text out)
• LLM integration: The full pipeline includes a finetuned Qwen3 model designed for word- and sentence-level decoding. This step is not entirely original, but it's still extremely important to help correct incorrect individual characters.

Congrats to the team on the launch - this is big for the entire ecosystem, and I am excited to see how many people are talking about it!