This week in AI, 13 to 17 October 2026: five releases and how to try them

Five things happened in AI between 13 and 17 October 2026 that are worth knowing about. One is a research result you can read and check yourself. Two you can try today with a paid Google plan, one you can try with a free sign-up, and one is a model you can download, if you have the hardware.

None of these is made by webridgeAI. Every detail below was checked on the company's own page on 17 October 2026, and each section links to it. Several news roundups this week listed releases we could not find on any company's page, so they are not here. Plans and prices change, so check the linked page before you pay for anything.

1. The five at a glance

2. Claude traced the NIST curve seeds

Most secure connections on the web can use a small set of elliptic curves that the US standards body NIST published around 2000. Each curve was made from a "seed", a 20-byte number that the NSA supplied in the 1990s without saying where it came from. That gap fed decades of suspicion that the curves might hide a back door.

On 16 October, Anthropic published the answer. An Anthropic employee asked an internal Claude model to look into it, and in under half an hour it found that most of the seeds were not made by a hash function, as people had long assumed. They came from lrand48, a common random number generator of the time, started with a Unix timestamp from 1997 or 1998. The P-256 seed, for example, traces back to 16 August 1997, 01:52:34 UTC. Claude rebuilt the generator's state from the published numbers, and the seeds it reproduced matched exactly.

What it means, in Anthropic's words, is that this is not a break and the curves are not weakened. If anything, it makes a hidden back door less plausible, because the seeds came from a simple process anyone can now repeat. There is nothing for you to do. Some base points and one binary-curve seed are still unexplained. Anthropic released the code to reproduce the seeds and gave the $36,864 bounty for solving it to the Internet Security Research Group.

3. Gemini 4 Argon in the Gemini app

Google announced Gemini 4 Argon, its newest model, on 30 September, and on 13 October it reached the Gemini app for Google AI Pro and Ultra subscribers. Google says it gives more direct answers, with less hedging and repetition, and reports its progress more often on long tasks. The same update put Google Antigravity's coding view inside the Gemini app for macOS.

In Google's own demo, a person uploads a photo of their daybed with a sketch, asks for "a 3d design app" to explore new legs for it, gets a working app with sliders, and 3D-prints the legs they pick. It is a demo, so expect to iterate, but it shows what to ask for: describe the tool you want, not just the answer.

How to try it: open the Gemini app or gemini.google.com, choose Argon in the model picker and describe the tool you want it to build. Developers can use it through the Gemini API. Google's announcement gives an introductory price of $2 per million input tokens and$10 per million output tokens, rising to $4 and$20 after the introductory period, and an output limit of up to 1 million tokens.

4. Muse Spark Open: a frontier model you can download

On 15 October, Meta released the weights for Muse Spark Open, which it calls the most capable US open-weight model to date. It has 816 billion parameters in total, of which 42 billion are used for each token, and a context window of 1 million tokens. It takes text, images, video and audio as input, and Meta says it is built for coding and long, multi-step work.

"Open weights" means you can download the model and run it on your own machines. It does not mean small. Apple ran it on Mac Studios with the M5 Ultra chip: at 4-bit precision the model needs about 493 GB, more than one 512 GB Mac Studio holds, so it is split across several. On four of them, Apple measured just over 36 tokens a second.

How to try it: most people should use a hosted version. Meta lists Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. The weights are on Hugging Face, and the model is released under the Muse Model License, which has limited commercial terms, so read it before you build a product on it.

5. Krea 3: photo edits in under eight seconds

Krea 3 is an image editing model. You give it a photo and say what to change ("make the lips cherry red", "add a white dove") and it changes only that, leaving the rest of the image as it was. Krea says an edit takes under eight seconds and costs about $0.06, and that it is four times faster than Google's Nano Banana Pro. Those are Krea's own figures, not our tests.

How to try it: sign up at krea.ai (there is a free sign-up), upload a photo and type one change. Start with something small, like a colour, so you can see what it leaves alone. How many Krea 3 edits each plan includes is on Krea's pricing page. Developers can call it through Krea's API.

6. Gemini tasks: busywork while you are away

The same 13 October update added tasks to the Gemini app. A switch in the main chat window moves Gemini from chatting to a task, and it keeps working when you leave your computer. Google says Gemini now connects to more than 40 apps, including Instacart, OpenTable, Adobe and Shopify, alongside Google's own apps.

In Google's demo, a small catering business asks Gemini to handle new inquiries. Gemini reads a request in Gmail, adds the client to a tracking sheet, builds a proposal deck in Google Slides and drafts the reply.

How to try it: with Google AI Pro or Ultra, open the Gemini app, switch to Task and describe one chore you repeat every week, with the apps it should use. Check its first few results before you let it send anything on your behalf.

7. Sources

All checked on 17 October 2026.

• Anthropic: Claude reveals ECDSA curve origins
• Google: Gemini 4 Argon in the Gemini app and the Gemini 4 Argon announcement
• Meta: Introducing Muse Spark Open
• Apple: Local and distributed LLM inference with MLX on Mac Studio
• Krea: Krea 3