Morning digest: OpenAI's raise, agent watchers, and Copilot

Morning digest for Tuesday, Oct 20. A money story and a security story lead, then two new tools for 3D and game assets, a science result from Anthropic, and a correction-style update on yesterday's Copilot item.

OpenAI is racing to raise about $30 billion before Anthropic goes public.

Reuters reports, citing two people familiar with it, that OpenAI wants investor commitments before Anthropic's expected IPO in mid-November, in a round that could value it above $1.4 trillion. SoftBank and Abu Dhabi's MGX are in talks to join. One source says OpenAI is telling investors it's at $50 billion in annualized revenue, expects $70 billion by year-end, and can keep lowering prices on its models. It was last valued at $852 billion in March. Separately, some Anthropic investors have pushed for an IPO valuation near $1.5 trillion instead of the $2 trillion it wanted. OpenAI and SoftBank declined to comment. What this means for you: nothing changes for ChatGPT or Claude users today. The part worth watching is OpenAI's pitch that it can cut model prices further, since that's the bit that would reach your bill. (Reuters via Business Standard)

Microsoft says "frontier models, regardless of their origin, cannot be trusted," and is selling tools to watch them.

In a Monday blog post, Microsoft security chief Hayete Gallot said AI agents inside companies should be treated like insiders who might go off-script, not because they're malicious but because they're unpredictable. Microsoft's answer is Guardian Agents, part of its Project Perception security system, which supervise what other agents do before they do it. TechCrunch reports Microsoft Defender can now spot up to 60 kinds of agents employees bring to work, including Claude Code, Codex and OpenClaw, and can limit their data access, adjust permissions per task, and require human approval for risky actions. TechCrunch also points out the awkward part: Microsoft has invested in both OpenAI and Anthropic. What this means for you: if you run Claude Code or Codex on a work laptop, expect your IT team to be able to see and fence it in. Solo builders can copy the idea for free: give each agent its own key, the smallest permissions that work, and an approval step for anything destructive. (Microsoft, TechCrunch)

Viggle's JST-2 builds a 3D scene in about a second and makes characters you can animate without rigging.

Viggle, the company behind the meme-video app, released JST-2 on Monday. It builds explorable 3D scenes from a prompt or images in about one second, which Viggle compares with 5 to 6 minutes for World Labs' Marble. It also makes 3D characters from a single image that are ready to animate with no separate rigging step, and turns text or video into full-body motion. In Viggle's own blind tests, people preferred its characters over Meshy 7.1 and Tripo H3.1, and its motion over Cartwheel and DeepMotion. You can try it free in the browser through Viggle's PINOC app. One catch for game work: scenes come out as 3D Gaussian splats, not ordinary meshes, and object collisions still use regular rule-based physics. Try it: the character and motion side is the part to test against your current tools. Compare it with the picks in our image-to-3D and AI mocap comparisons, because the benchmark numbers are Viggle's own. (Viggle)

Ludo.ai's new command-line tool lets Claude Code make your game's assets.

Ludo.ai released the Ludo CLI on Monday. It lets you, or a coding agent like Claude Code, generate sprites, animations, textures, 3D models, video, sound effects, music and voices from the terminal, saved straight into your project folder. You can preview credit costs and set spending limits per project. It needs a Pro plan ($35 a month billed yearly, $420) or Studio ($300 a month billed yearly). The $15 Indie plan doesn't include it, and every CLI call uses credits even for things that are unlimited in the web app. What to weigh: if you already build with Claude Code, it's a simple way to let it fill in placeholder art while it writes the game, though we haven't tested output quality yet. If you only make assets by hand now and then, the web app is cheaper. (GamesBeat, Ludo pricing)

Claude designed proteins that latch onto scorpion toxins, and 17 of them worked in the lab.

Anthropic said this morning that an internal research model, working mostly on its own for 64 hours with four follow-up prompts, generated more than 318,000 protein designs aimed at scorpion neurotoxins. Two outside labs tested 211 of them, and 17 bound to their target toxin. Anthropic is clear that it hasn't shown they neutralize venom, so this isn't an antivenom. It's releasing all 211 designs and their data. Scorpion stings hit more than 1.2 million people a year and kill more than 3,250, mostly children. Why it matters: this is a neglected disease with little money behind it, and the slow, costly step of finding binders took days. Turning them into a safe, cheap drug is still the hard part. (Anthropic)

Update on yesterday's Copilot retirements: GitHub now points to newer replacements.

GitHub's day-of notice on Monday suggests GPT-6.1 Sol in place of GPT-5.5 and GPT-5.4, GPT-6 Luna in place of GPT-5.4 mini and GPT-5 mini, and Grok 4.7 in place of Grok 4.5. That's newer than the September notice we quoted (GPT-5.6 Sol, GPT-5.6 Luna, Grok 4.6). Gemini 3.8 Flash is unchanged. It also says Enterprise admins may need to turn the new models on in their model policies, and GitHub released an API for checking and changing those policies at the enterprise, team, app and organization level. Try it: if a workflow is still pinned to a retired model, move it to the newer names. (GitHub, Oct 19 notice, model policy API)

Meta shows the Ohio supercomputer that trained Muse.

Meta's engineering blog gave a first detailed look at Prometheus in New Albany, Ohio. Phase one is finished and trained the Muse models, including Muse Wonder. Meta plans to grow it to about a million GPUs. Training interruptions peaked at almost 1,000 a day per 8,000 GPUs in mid-2025, and Meta says it cut that by 99% in a year. What to watch: Meta says big training runs can swing power use by tens of megawatts in seconds, which is why power, not chips, keeps showing up as the limit. (Meta Engineering)

Gemini Notebook now lets you check the plan before it makes slides or a video.

Ask for a slide deck or a Video Overview and Notebook suggests topics from your sources first. Turn on the new Review toggle to read the outline or "producer notes," ask for changes like an executive summary up front or a different scene order, then hit Proceed. It's rolling out to all Gemini Notebook users over the next week. Try it: turn Review on for anything long, so you fix the plan once instead of re-rendering. (Google)