The UK AI Safety Institute found that OpenAI's GPT-5.5 is almost as capable as Anthropic's Mythos model at executing multi-step cyberattacks, eroding the expected lead time for defenders.
OpenAI’s GPT-5.5 Instant model is now the default for ChatGPT, bringing deeper personalization and a new 'Memory sources' feature that shows you exactly which past conversations influenced its responses.
OpenAI’s GPT-5.5 Instant model is now the default for ChatGPT, bringing deeper personalization and a new 'Memory sources' feature that shows you exactly which past conversations influenced its responses.
Codex's new Pets feature isn't just decoration — it's a persistent status indicator that tells you what your agent is doing even when you're in another app. You can even create custom pets based on your projects.
OpenAI's Codex CLI now has a /goal command that keeps running endlessly until the goal is met, eliminating the need for manual scripting or bypass flags.
OpenAI Codex adds an office mode alongside coding, signaling a dual-track trend for AI agents. Anthropic and ByteDance follow suit.
From GPT-5.4 onward, Codex is no longer a separate model line. OpenAI's head of developer experience confirms the merge, and GPT-5.5 brings enhancements to agentic coding and computer use.
Hold the Fn key and speak into any macOS app — Codex's new global dictation transcribes at the cursor, no app switching required.
OpenAI's Codex can now autonomously test and debug front-end code by opening a built-in browser, clicking through the UI, and using visual and network logs — no manual switching needed.
Andon Labs pitted GPT-5.5 against Opus 4.7 in a vending machine simulation. GPT-5.5 won without cheating, while Opus's lies and collusion barely helped — and sometimes hurt.
