Monday, April 13, 2026

Meta’s New Model Ups the Stakes

Meta’s New Model Ups the Stakes

Today’s Overview

Good morning, Meta, OpenAI, and Anthropic all made big moves at the top of the stack, while a few sharper research and product updates filled in the picture. The most interesting thread is how fast AI systems are shifting from chat to orchestration, whether that means coding, security, or full multi-agent workflows. Let’s dive in.

Top Stories

Meta launches Muse Spark

Meta released Muse Spark, the first model from its Superintelligence Labs, and put it into Meta AI with a private API preview. The compact multimodal system adds tool use, visual chain-of-thought, multi-agent orchestration, and a Contemplating Mode for parallel reasoning. It is already live and Meta says open-source plans are still ahead.

  • It is the first model from Meta’s Superintelligence Labs after Llama 4 fell short.
  • The system adds tool use and multi-agent orchestration along with a Contemplating Mode for parallel reasoning.
  • Meta says the model is already live in Meta AI with a private API preview and future open-source plans.

OpenAI cuts Pro to $100

OpenAI launched a $100/month ChatGPT Pro tier for heavy Codex users, with roughly 5x the usage limits of Plus for agentic coding, cloud tasks, and code reviews. It also clarified that its invite-only cyber offering is a standalone defensive product. On the business side, the company said it will reserve IPO shares for retail investors as it prepares to look more like a public company.

  • The new plan targets heavy Codex users with roughly 5x the usage limits of the $20 Plus tier.
  • OpenAI also clarified that Trusted Access for Cyber is a standalone defensive product.
  • The company said it will reserve IPO shares for retail investors as it prepares to act more like a public company.

Anthropic unveils Mythos Preview

Anthropic said Claude Mythos Preview is its most powerful model yet, but it will not be generally released. Instead, the company is channeling the work into Project Glasswing, a cybersecurity initiative for about 50 organizations aimed at finding and patching vulnerabilities before attackers can exploit them. Anthropic says the safeguards developed there will carry into a future Claude Opus model.

  • Mythos Preview reportedly found thousands of zero-day vulnerabilities across major operating systems and browsers.
  • Anthropic plans to support the effort with up to $100 million in usage credits and $4 million in direct donations.
  • The model will not be generally available; the security work will ship with a future Claude Opus model instead.

Research & Analysis

Oxford AI spots heart failure early

Researchers at the University of Oxford introduced an AI system that reads changes in the texture of fat around the heart from routine CT scans. It reportedly reached 86% accuracy across 72,000 patients and can flag high-risk cases years before symptoms appear. Oxford is working with regulators to bring the tool to National Health Service hospitals and wants to extend it to all chest CT scans within months.

  • The system reads changes in fat around the heart from routine CT scans.
  • It reportedly achieved 86% accuracy across 72,000 patients.
  • Oxford says the tool could reach NHS hospitals and expand to all chest CT scans within months.

AI surfaces GLP-1 side effects

Penn researchers used AI to analyze more than 400,000 Reddit posts about Ozempic and Mounjaro through computational social listening. The work mapped posts from 67,000 users to standardized medical terms and found side effects that are not reflected in current drug labels. Fatigue stood out as one of the most common complaints, while chills, and hot flashes also appeared.

  • The team analyzed more than 400,000 Reddit posts about Ozempic and Mounjaro.
  • The models mapped discussion from 67,000 users to standardized medical terms.
  • The study found side effects not reflected in labels, including fatigue along with chills and hot flashes.

Research-first coding agents perform better

A coding-agent workflow that reads papers and studies competing projects before writing code produced better optimizations than code-only prompting. The takeaway is simple: adding a research phase helps agents form stronger hypotheses before they touch the implementation. It is a small change with a big impact on quality.

  • Agents that start with reading papers form stronger hypotheses before they code.
  • The workflow beats code-only prompting for optimization work.
  • The main shift is adding a research phase before implementation begins.

NVIDIA’s Sol-RL trims rollout costs

NVIDIA’s Sol-RL uses a two-stage framework that separates exploration from training. It combines FP4 rollouts with BF16 policy updates to cut compute costs while improving alignment and speeding convergence. The result is a more efficient approach to diffusion model post-training.

  • The system separates exploration and training into two stages.
  • It uses FP4 rollouts to generate large candidate sets.
  • Policy updates happen in BF16 to reduce compute costs while improving alignment.

Trending AI Tools

  • Perplexity Plaid integration Connects bank, credit, loan, and brokerage data so Computer can build budgets, net worth trackers, and debt payoff plans.

  • Gemini app updates Adds real-time simulations, 3D models, persistent Notebooks, and step-by-step coding help in Colab.

  • Claude Managed Agents Brings cloud orchestration, sandboxing, persistent state, and new enterprise controls to Claude.

Quick Hits

Keep reading for free

Enter your email. If you're already subscribed, we'll send a sign-in code. If not, you'll subscribe in the next step.

Free access. Subscribe once, then use the same email on future issues.

Free to read. Subscription just unlocks the full issue.