🤖 GPT-6.2 Astra surfaces in Codex ahead of launch 🤖
A staged rollout hints OpenAI's next coding model is close.

GPT-6.2 Astra surfaces in Codex ahead of launch, according to TestingCatalog, which spotted the GPT-6.2-Astra identifier inside Codex usage breakdowns. The appearance points to an A/B test or staged rollout, a common step before a public release, meaning some developers may already be using the model without knowing it. The report also claims OpenAI skipped GPT-6.1 Astra entirely over concerns about deception and scope, a rare signal that safety reviews are deciding which versions actually ship. The move adds competitive pressure on rival coding agents like Claude Code. For teams building on Codex, it is worth tracking output quality and behavior over the coming weeks, since no official launch date has been confirmed yet.

💳️ Anthropic renames Claude plans, frees Ultra limits 💳️
New pricing tiers bring simpler names and fewer Ultra caps.

Anthropic renames Claude plans, frees Ultra limits in a quiet but meaningful update to its subscription lineup. The plan formerly called Max 5x is now simply Max, while Max 20x becomes Ultra, replacing multiplier-based labels with names that are easier to compare at a glance. The bigger change sits under the hood: verified Ultra subscribers no longer face 5-hour usage limits, the rolling caps that heavy users of Claude Code and long agent sessions have often run into. Removing those caps is a clear pricing play for power users. It also raises the stakes in the premium subscription race with OpenAI and Google, where usage ceilings increasingly decide which top tier professionals choose to pay for each month.

🔐 GitHub enclaves lock down agentic workflows 🔐
A new security primitive tightens integration across multi-repo agents.

GitHub enclaves lock down agentic workflows by introducing a new security primitive for cross-repository GitHub Agentic Workflows, published by the GitHub Next research team. As AI agents increasingly operate across several repositories at once, they can pass context and code between each other in ways that are hard to audit. Enclaves limit what agents share with each other across repos, creating defined boundaries so that an agent working on one project cannot freely leak information into another. For engineering teams, this lowers the compliance risk of letting autonomous agents touch sensitive codebases. It reflects how quickly multi-agent automation is moving from experiment toward everyday development work.

INTERESTING TO KNOW

📱 Claude threads cut agent latency by 35% 📱

Claude threads cut agent latency by 35% thanks to a speed-focused developer update now in public beta: with message threads, agents send only new messages instead of resending the full conversation history on every call. For long-context requests between 500K and 1M tokens, Anthropic reports a ~35% drop in median time to first token, a direct responsiveness gain for teams running long-lived agents in production.

🦾 Meta builds Max humanoid for household chores 🦾

Meta builds Max humanoid for household chores, and the speed is notable: the company reportedly assembled the five-foot robot in just three months with Chinese robotics firm Fourier, running on Nvidia Jetson Thor. A stair-climbing Max 2 is already in development, a signal that Meta is pushing AI beyond screens and glasses into physical homes, where Tesla and Figure are already competing.