If Anthropic is following the path of integrating Claude Code into its Work desktop app (finally adding parallel sessions: https://claude.com/blog/claude-code-desktop-redesign), OpenAI is coming from a different angle: this week they updated the Codex coding app and added computer control features. Different paths — same result.

Codex as a Superapp
https://openai.com/index/codex-for-almost-everything/
On macOS, Codex can now see the screen, move its own cursor, click, type text, open any application, and work in the background. Across all platforms, there is a built-in browser, image generation, memory (remembers preferences and previous actions — not yet in EU/UK), and over 90 plugins and integrations.

https://www.youtube.com/watch?v=sdNoaztocs0

While Codex has introduced a very "Cursor-like" pleasant feature — where you can simply click on any element (button, block, text, image) in a generated website to immediately add it to the prompt as a reference — the general trend of both companies (Anthropic and OpenAI) expanding their product audiences is slightly concerning for programmers.

Discussion
https://news.ycombinator.com/item?id=47796469
Many see this as a revolution for ordinary people (non-programmers): agents will be able to create personal UIs, automate business processes, replace entire programs, and radically increase productivity. At the same time, programmers are wary — security and privacy are still neglected: full agent access turns a computer into a "hostile device" where even a .txt file becomes an attack vector.

ChatGPT Pro for $100/mo
https://help.openai.com/en/articles/9793128-about-chatgpt-pro-tiers
At the beginning of April, the Codex token promotion ended; now, a free account can only run about two simple tasks before hitting the weekly limit. The $20 Plus plan also doesn't offer much headroom now, as the weekly limit is only suitable for light work (1-2 hours a day). That’s why, as of April 9th, an intermediate option between Plus and the $200 Pro was added.

The new $100 Pro tier has 5× higher limits than Plus and provides access to GPT-5.4 Pro and GPT-5.3 Instant. There is also a promotion until May 31, 2026, offering double tokens.

This is a direct response to Anthropic, which has Claude Max for $100.

Opus 4.7
https://www.anthropic.com/news/claude-opus-4-7
Claude Opus has updated from 4.6 to 4.7 — same features, but even better on benchmarks. They added "adaptive thinking": the model decides for itself how much to "think" before responding and hides the internal reasoning (it no longer shows the full chain of thought by default).

Discussion
https://news.ycombinator.com/item?id=47793411
The model has become stronger, especially in coding and long contexts. However, it is becoming less debuggable. It is now impossible to properly disable adaptive thinking, which makes Claude Code even worse to use; one has to jump through hoops with commands like /effort xhigh, CLAUDE_CODE_DISABLE_1M_CONTEXT=1, "display": "summarized", etc., just to understand what the model is generating.

Anthropic makes cool models, but the programming tools around them are getting worse.

#newllmmodel #codex #claudecode