OpenAI Codex Update: First Step Toward a Superapp

OpenAI's Big Move: Codex Evolves Toward a Superapp

Good morning, {{ first_name | AI enthusiasts }}. OpenAI has been teasing a superapp for months. Today it shipped a major first piece: a significant Codex update that expands the tool well beyond a coding assistant.

Codex now includes background computer use, an in-app browser, parallel agents, image generation, and session memory (preview). These additions make Codex feel more like ChatGPT + Atlas + Codex rolled into one — a clear step toward an all-in-one platform.

OpenAI calls this "building the super app out in the open." Codex now supports long-running automations, multiple agents working in parallel, and native mockup generation.

Key features in the update

• Background computer use: Codex can operate Mac apps on its own, even when apps don't expose APIs.
• Parallel agents: Multiple agents can run at once to handle separate tasks or collaborate.
• Memory (preview): Preferences and session context persist across conversations.
• Automations: Long-running tasks can resume days later.
• Atlas-powered in-app browser: Developers can annotate pages to guide Codex actions.
• Inline image generation: gpt-image-1.5 creates mockups without switching apps.

Codex has grown quickly, reaching 3 million weekly users with 70% month-over-month growth. This expansion puts OpenAI in a more direct competitive stance with Anthropic's agent-focused offerings.

Anthropic's Opus 4.7: Strong on agentic coding

Anthropic released Claude Opus 4.7, a public model that leads on agentic coding benchmarks compared with some rivals. It still trails Anthropic's gated Mythos Preview, however.

Highlights

• Opus 4.7 jumps to 64.3% on the SWE-bench Pro coding benchmark (from 53.4% in 4.6).
• Mythos Preview remains ahead at 77.8%.
• New defaults and tools: a mid-tier 'xhigh' effort for Claude Code and an /ultrareview command to flag bugs and design issues.
• Some users reported degraded performance on 4.6; reactions to 4.7 are mixed despite benchmark gains.

Anthropic appears to run two tracks: a fast public cadence and a gated frontier track (Mythos) for partners. That split helps test cutting-edge models but means public access can lag the real frontier.

Run an LLM locally with Ollama

You can run a real large language model on your laptop for free using Ollama. No subscription, no account, and no data leaves your machine.

Quick setup

1. Download the installer from ollama.com/download for Mac, Linux, or Windows and install it.
2. Open the app, choose New Chat, and select a lightweight model like gemma3 (~3GB) suitable for 8GB RAM laptops.
3. Wait for the model to download and then type a prompt. The model runs locally — try airplane mode to confirm.

Pro tip: Use Ollama's API to connect the local model to web tools, or point a coding agent like Claude Code at it to offload tasks.

OpenAI's science model: GPT-Rosalind

OpenAI launched GPT-Rosalind, its first life-sciences domain model tailored for drug discovery and biological research. Rosalind reads papers, queries lab databases, designs experiments, and generates hypotheses.

• Rosalind outperforms GPT-5.4 on several science-focused benchmarks, including biochemistry and experiment design.
• On a blind RNA test from Dyno Therapeutics, Rosalind scored better than 95% of human scientists in prediction tasks.
• It's currently available to qualifying enterprise users; organizations such as Amgen, Moderna, and the Allen Institute are testing it.

Why this matters: Purpose-built models like Rosalind show that while flagship models aim to be generalists, industry-leading advances often require domain specialization — especially in fields like drug discovery and cybersecurity.

Trending AI tools and industry news

• Claude Opus 4.7 — Anthropic's strong agentic coding model.
• Windsurf 2.0 — An agentic IDE with a command center for fleets of agents.
• Codex — OpenAI's agent now with computer use, in-app browser, and image tools.
• HY-World 2.0 — Tencent's open-source world model for editable 3D scenes.

Other updates

• Perplexity released a Max-tier Mac app, Personal Computer, to run agents across many frontier models and pilot a browser 24/7.
• Windsurf added an Agent Command Center for coordinating cloud and local agents.
• Tencent open-sourced HY-World 2.0 for physics-aware 3D scene generation.
• Reports suggest the U.S. government may grant some agencies access to Anthropic's Mythos AI despite ongoing legal and policy friction.
• Alibaba introduced Happy Oyster in beta, a world model that builds interactive 3D environments from multimodal inputs.

Community workflow spotlight

Each issue we share a reader's practical AI workflow. Today's comes from Jerry G. in Gig Harbor, WA:

"I'm a 73-year-old author and screenwriter. I have been using Claude to build my last two books. Claude helped design the book covers, formatted the interior, and even suggested titles. Now that they are printed, Claude helps me market them to book clubs, libraries, and social media sites. Claude helped me build a Substack following and post bi-weekly new stories on Instagram and BlueSky. I'm a Claude convert and won't write any more books with his assistance."

How do you use AI? Tell us here.

Why this matters for gaming news and broader industries

Although this update centers on coding and scientific models, the move toward superapps and parallel agents will affect many spaces — including gaming news, game development tools, and interactive content creation. Developers can use multi-agent workflows to prototype game mechanics, generate art mockups in-app, and automate long-running testing tasks across platforms.

Expect faster iteration cycles in game development, more powerful tooling for indie studios, and richer interactive experiences powered by domain-specific and agent-driven AI.

See you soon,

Rowan, Joey, Zach, Shubham, and Jennifer — the humans behind The Rundown