OpenAI Lets ChatGPT Pro Run Several Agents at Once

On October 16, 2026, OpenAI upgraded GPT-6 Astra Pro, the model in ChatGPT Pro, OpenAI's most expensive plan. Several agents can now take different parts of one task and work on them at the same time. You can follow each one. When an agent finishes, you open its work and check it before you use it. Pro subscribers get this in the Chat tab. OpenAI announced the change only on X, and it has not said how many agents can run at once or how a task gets split between them. OpenAI's post ↗

Update: Reflection AI Releases Beam for Download

Update: On October 16, 2026, Reflection AI put Beam, its model for coding and agent work, on Hugging Face. The license is Apache 2.0, so businesses can use it in their own products. On October 6, when News Desk first covered Beam, you could only join a waitlist. Reflection says Beam scores 80.9 on SWE-bench Verified, a test where models fix real software bugs. Nobody has checked that number independently yet. The full model takes about 1TB of memory, so most people will use it through a hosted service, not on their own computer. Reflection's announcement ↗

Anthropic Explains How It Reviews Code Written by Claude

On October 16, 2026, Anthropic published a post about how it checks the code Claude writes for Claude Code, its coding tool. A second copy of Claude reviews each change from a fresh start, with no memory of how the code was written. Several reviewers look for bugs at once and then vote on what they found. Next, a classifier decides if zero, one or two people also need to look. A change that only touches documentation needs no person at all. Teams that already let AI write their code can copy the method. The post also explains how to set it up. Anthropic's post ↗

Hark's AI Ordered Thousands of Free Pizzas Across the US

Hark, an AI assistant app, gave away free pizza across the US on Friday night, October 16, 2026, for the launch of Hark Pro. People asked the app for a pizza. Then Handoff, Hark's model for using websites, did the whole order: it logged in, worked through the site and paid. Brett Adcock of Hark says usage reached record levels and a lot broke under the load. Most AI tools that use websites are still demos, so this was a rare public test, with real orders and real payments, at a scale anyone could see. Brett Adcock's thread ↗

Four Gemini 4 Argon Agents Beat Minecraft in 27 Hours

Google DeepMind put four copies of Gemini 4 Argon, Google's newest model, into one Minecraft world. On October 16, 2026, DeepMind said they had beaten the Ender Dragon, the game's final boss. The agents saw no screen. They worked with code and text only, split up the jobs over chat and fixed each other's errors with no human help. The log lists 67 drownings. This is a research demo, not a product. Today's ChatGPT update uses the same idea of several agents splitting one long job. DeepMind's post ↗

Elorian Releases Glimpse, a Model That Works With Images

Elorian, a startup in Palo Alto, released its first model, Glimpse, on October 16, 2026. The company raised $300 million before it had a product. While it works out an answer, Glimpse can zoom into part of a picture, pull single frames from a video and run small programs. Elorian says Glimpse reads scientific charts better than GPT-6 Astra, but no independent test has checked this. Developers can use Glimpse now. The chat app has a waitlist. Elorian's announcement ↗

Anthropic's Claude Code Plugin Directory Gains 2,328 Stars

This week, anthropics/claude-plugins-official gained 2,328 stars on GitHub. Anthropic runs this directory. It lists plugins, which are add-ons that give Claude Code new commands and skills. If you use Claude Code, start here to find add-ons. Anthropic chooses what goes on the list. the directory ↗