The AI Roundup: Week 36

This episode is AI-generated, voice and picture included. MiniMax lifts its purchase ceiling at Alibaba Cloud by 220 percent to 1.2 billion dollars. Compute bills are growing faster than planned.

MiniMax is a Chinese AI developer, and the company has raised its purchase ceiling at Alibaba Cloud. The three-year frame rises by 220 percent to 1.2 billion dollars. This is a ceiling, not money already spent. MiniMax gives the reason itself: training and running its own models consume more compute than it budgeted for. This matters because it puts a number on something contracts usually hide. Anyone building models does not budget compute once, but writes it upward again and again. And anyone buying those models later carries that curve too.

A paper from the UIST conference describes a system called Co-Annotator. It records where experienced retina specialists look while reading OCT scans, and it captures their dictation at the same time. Two assistance features for the reading task are built from both. According to the paper, correct diagnoses per minute rise by 40 percent, and the time spent editing report comments falls by 67 percent. This matters because the model here is not the text a doctor writes, but where a doctor looks. Experience becomes recordable. The figures come from a single source so far. We report them as a claim by the authors, not as a confirmed result.

Texas is putting no further state money into the camera maker Flock for now. Governor Greg Abbott froze the spending on its AI surveillance cameras. The freeze came shortly before a report on purchases worth more than 30 million dollars. New cameras are therefore not being paid for out of state funds until further notice. This matters because what slowed this down was neither a court nor a law, but a budget decision. Surveillance technology spreads through procurement, and procurement is also where it can be halted. So if you want to know how far such technology reaches, look at the invoices.

OpenGrok is a community project under an MIT license. It puts a model picker in front of Grok Bot and routes requests to outside providers, among them GLM, Claude, Gemini, DeepSeek, or a model running locally on your own machine. Six providers are marked working, and three are still waiting on a capture of the data connection. The project says it has no affiliation with xAI. This matters because the interface and the model turn out to be easy to separate. Anyone who likes the bot but wants a different model behind it does not need the provider to agree. For everyone else it is a hint at how interchangeable the models behind familiar interfaces have become.

Anthropic has introduced two new models: Claude Fable 5.1 and Claude Mythos 5.1. It gives prices of 10 and 50 dollars per million tokens, with 10 dollars covering one million input tokens for Fable 5.1. On the Terminal-Bench 4.0 test Anthropic reports 55.8 percent. Fable 5.1 runs on the Claude API and on three cloud platforms. Mythos 5.1 goes only to vetted organizations in the United States. This matters because two paths now sit side by side. One model is broadly purchasable, the other is open only to a vetted circle. Anyone working outside the United States can use the one and not the other.

Runway is showing a research model called Solaris. It does not build an interface out of code, but generates it as video at 720p, one frame at a time, while you use it. It responds to clicks, to dragging and to voice. There is no launch date so far. This matters because software here is no longer executed, but generated. What appears on the screen comes into being at the moment you touch it. As long as no date is named, this stays research and not a product you can plan around.

CrowdStrike and NVIDIA used the Fal.Con 2026 conference in Las Vegas to launch a system called SafeMind. It is an agentic defense: attack and defense models keep training each other. The defensive models are post-trained from the open Nemotron family from NVIDIA and run natively inside the Falcon platform. This matters because defense here does not wait for a known signature, but practices on its own. Anyone deploying such systems moves part of the security work to models that keep learning while running. That changes who you need on a security team as well.

Google DeepMind has rebuilt video analysis in Gemini. Since September 1, 2026, the model itself decides which passage of a video to open, at what frame rate to sample it, and whether to use frames, audio or the transcript. For long videos the company reports up to 88 percent fewer tokens. This matters because the cost of video hangs on the number of frames sampled. Anyone searching longer material has so far paid for every single one of those frames. The saving, however, comes from the company itself and has not been measured by us.

In a post of its own, OpenAI presents three companies that have built agents into their workflows: Basis, Clay, and Exa Labs. The work named is onboarding, account management, and developer integrations. We could not retrieve the full text of the post, so we have no impact figures. This matters because the provider is picking the examples here. What remains is the direction: agents no longer sit beside the work, they sit inside the workflows. How well they actually perform there is something we cannot verify at this point.

The sources for every story sit below the matching articles on ai-in-life.com. The English versions are there as well. See you next week.