It was known that the model update announcements from the two top companies were scheduled for the same time on February 5, 2026, but then Anthropic went live 15 minutes early. However, with OpenAI's announcement, their model became available only within Codex, with no API access. This prevented third-party projects (like Cursor or Cline) from offering immediate access to it.
https://www.youtube.com/watch?v=9f2egsZZjnw
Update to Claude Opus 4.6
https://www.anthropic.com/news/claude-opus-4-6
Anthropic has improved upon Opus 4.5. It features enhanced skills in planning, autonomous operation, code review, document handling, and online search. The beta version includes a 1M token context window and automatic summarization of older context for lengthy tasks (Context Compaction). The key feature is the ability to execute longer and more complex tasks autonomously.
https://code.claude.com/docs/en/agent-teams
Claude Code has added Agent Teams for the autonomous coordination of multiple agents. Unlike sub-agents, which operate within a single session where interaction is restricted to the main agent, here you can interact directly with individual "team members" without going through the team lead.
HN Discussion
https://news.ycombinator.com/item?id=46902223
Skepticism outweighs enthusiasm. Many users fail to notice a significant difference between 4.5 and 4.6, with some even noting, "10x more expensive than Sonnet, but no difference." The general consensus is that "all models have their flaws." There is widespread criticism of Claude Code for its slowness, high memory consumption, and the use of React for the terminal.
Update to GPT-5.3-Codex
https://openai.com/index/introducing-gpt-5-3-codex/
An improvement over GPT-5.2-Codex. This is a specialized model for generating code for complex projects and automation. It aims to be 25% faster than 5.2-Codex while maintaining the same accuracy.
The main focus of the announcement is Interactive Collaboration. You can "steer" mid-execution—meaning you can re-prompt the model without stopping it, and it will immediately shift its strategy. This contrasts with Opus 4.6, which attempts to work autonomously for extended periods with minimal human intervention.
Codex as an App
https://openai.com/index/introducing-the-codex-app/
In addition to the CLI and IDE extension, there will now be a standalone app under this name. It is built on Electron, though only the Mac ARM version was available at launch, with a waitlist for other platforms. This is another attempt to create an agent "control center," similar to what exists in Cursor and Antigravity. This one seems successful.
https://www.youtube.com/watch?v=ICYbOfW5RoQ
It is a graphical interface (GUI) for the Codex CLI that allows for managing multiple projects, agents, and conversations in a single window. Features include quick switching between projects and chats, voice control, open in IDE, automatic builds, diff-view, terminal.
HN Discussion
https://news.ycombinator.com/item?id=46902638
Users highlight the diverging strategies chosen by the top players. With Codex, it's "Steering mid-execution"—the ability to control the process while it runs. The human stays "in the loop." Faster recovery from errors. Better performance with backend and "hard" tasks. Claude focuses on increased autonomy for agent swarms and long-duration tasks, but users note that the "Fire and forget" approach often leads to chaos and poor-quality code.
I believe OpenAI has made a series of good decisions regarding code generation for professional programmers, as opposed to "vibe-coders" or prototypers. The latter group is better suited for Opus 4.6, which will devour a lot of tokens as a swarm in Claude Code, but eventually generate a working version.
I like that, as of recently, Codex models have started writing back how they understood me after my request, and at every step, they report exactly what they are about to do. Generation can be quickly stopped if a misunderstanding occurs, allowing me to add new instructions and clarifications. Judging by the video, I see that in the new Codex app, the code being edited is hidden by default, showing only these text messages.
Furthermore, while working with the CLI, I built my own web app to manage all my chats across all my projects, because doing so from within the CLI is very inconvenient. The new Codex app, judging by the video, does exactly that—I'll be waiting for the Windows version.
#newllmmodel #codex #claudecode