GPT‑5.6, the new ChatGPT super app, and a massive week for AI

This has been one of the biggest weeks for AI in a long time. OpenAI pushed out a major new flagship model, rolled its tools into a single “super app,” and launched a real-time voice mode. At the same time, xAI, Meta, Anthropic, Google, and ByteDance all dropped notable updates of their own.

GPT‑5.6: a flagship model that feels like a full generation leap

OpenAI’s new GPT‑5.6 family is the headline of the week. Despite the incremental version number, it feels much closer to a GPT‑6‑level jump than a minor 5.5 update.

Previous steps from 5.3 → 5.4 → 5.5 were solid but incremental. In contrast, 5.6 dramatically improves reasoning, coding, and agent-style behavior while also being faster and cheaper for many workloads.

How strong is GPT‑5.6 on benchmarks?

On practical, task-focused benchmarks, GPT‑5.6 is now near the very top:

GDP‑Val (economically useful tasks): GPT‑5.6 Soul (the highest tier) beats almost every model except Anthropic’s Claude Fable, and even there the gap is small.

Browser and computer use: On benchmarks that test how well a model can control a browser and operate a computer, GPT‑5.6 outperforms almost everything, including Mythos and Fable.

GPQA (hard, proof-style questions): GPT‑5.6 ties Mythos on this notoriously difficult benchmark, showing strong high-end reasoning.

Buybench (a playful but telling “real work” benchmark): GPT‑5.6 sweeps the top four spots. GPT‑5.6 Soul Pro is #1, followed by 5.6 Terra Pro, 5.6 Terra, and 5.6 Soul. Previous Luna models don’t even crack the top 10.

Speed and price vs GPT‑5.5

One standout example: GPT‑5.6 Soul Pro generated a complex Buybench artifact using 45,000 tokens in about a minute for roughly $0.77. GPT‑5.5 Pro needed six minutes and about $4 to produce a similar result. That’s a huge jump in both speed and cost-efficiency.

What it can actually build

Benchmarks are nice, but the real story is what GPT‑5.6 can create:

Games and interactive apps: With a single prompt, GPT‑5.6 built a polished 3D shooter-style game (a Mega Bunk–style clone) using 3D graphics and smooth camera controls. On the first try, it produced better visuals and controls than an earlier attempt with Claude Fable, which needed extra prompting to reach the same level.

Beautiful websites and slide decks: GPT‑5.6 can generate full, interactive websites and slide decks that feel more like creative experiences than basic templates. One example site it built featured multiple surreal environments, interactive rain and particle effects, and dynamic transitions—all from a single prompt asking it to “build a website that will really impress me.”

How GPT‑5.6 compares to Claude Fable

Claude Fable and GPT‑5.6 are now the two standout frontier models. In earlier coverage of GPT‑5.5 and Fable, they were already close; GPT‑5.6 pushes OpenAI back to the front on many coding and agentic benchmarks. If you want a deeper comparison of the 5.x series and the broader model landscape, check out this breakdown of GPT‑5.5 vs DeepSeek V4 and other top models.

In practice, GPT‑5.6 feels like the best “daily driver” model right now: you can dump huge specs into it, tell it to build, walk away, and come back to something that’s 90–98% of the way there. Fable still seems slightly better at deep architectural reasoning and long-term planning, but GPT‑5.6 is often more thorough in implementation and even caught subtle security issues (like API key exposure) in code originally generated by Fable.

The new ChatGPT super app: work + code + browser + sites

Alongside GPT‑5.6, OpenAI launched a unified ChatGPT app that merges what used to be separate tools (like the Codeex app and Atlas browser) into a single “super app.” This is positioned as the main way to access OpenAI’s models going forward.

Work mode vs Codeex mode

The app has two main modes:

Codeex mode: Optimized for development. You get branches, local project folders, a terminal, code review tools, and direct integration with your filesystem. It’s meant for building and debugging software.

Work mode: Optimized for getting things done. Think of it as your AI personal assistant. You give it goals in plain language and it decides whether to answer directly, call tools, write code, browse the web, or orchestrate workflows across your apps.

The UI doesn’t change dramatically between modes, which can be confusing at first, but the intent is clear: Codeex is for hands-on coding, Work is for broader productivity and automation.

Integrated browser, files, and tools

In Work mode, the right sidebar gives you a built-in browser (replacing Atlas) and file access. In Codeex mode, that sidebar adds a terminal and review tools on top of browser and files. Everything runs inside the ChatGPT app, so you can browse, code, and test without constantly switching windows.

The app also supports a wide range of integrations—Gmail, Google Drive, Slack, Granola (for meeting notes), your local repositories, and more. This is what enables it to act as a real personal assistant that understands your entire digital workspace.

Model slider and advanced options

Inside the app, you can pick specific models (GPT‑5.6 Soul, Terra, Luna, or older 5.5 variants) or use a simple slider:

Low end: 5.6 Terra Light (fastest and cheapest).

High end: 5.6 Soul Ultra (most capable, most expensive).

Turn off “advanced mode” and the slider abstracts away model names entirely, letting you trade off speed vs intelligence without worrying about which variant you’re on.

ChatGPT Sites: prompt-to-production web hosting

One of the most powerful new features in the app is Sites, a built-in website builder and host.

From the app, you can click the “Sites” button, describe what you want, and ChatGPT will:

• Generate the full site (front end and any needed backend logic).
• Handle hosting under a chatgpt.site domain.
• Manage infrastructure details like databases, CDNs, and deployments for you.

In practice, this means you can build something like a 3D browser game or a complex interactive art site, then move it into Sites with a single command (e.g., “build this game on Sites”). After some build time, you get a live URL you can share, plus simple controls to make it public or private.

This is very similar in spirit to other “agents that build and host your apps” platforms. If you’re interested in that broader trend, there’s a good overview in this article on Abacus AI’s agent-based app hosting.

A true AI personal assistant inside ChatGPT Work

ChatGPT Work mode is designed to be more than just a chat window—it’s meant to be an actual assistant that understands your life and work.

Because it can connect to your Gmail, calendar, Slack, Google Drive, Granola meeting notes, your second brain (e.g., Obsidian), and your local projects, it can build a surprisingly detailed picture of what you’re doing and where you’re stuck.

Deep context and “control tower” dashboards

With permission, the assistant can scan:

• Recent emails and calendar events.
• Slack DMs, mentions, and threads.
• Meeting notes in Granola.
• Recently edited Google Docs and Drive files.
• Your second brain notes and journals.
• Local repositories and projects.

From there, it can propose a prioritized list of what you should focus on and what it can handle for you. One example setup had it build a “daily control tower” dashboard that tracks:

• What needs attention today vs what can wait.
• Tasks blocked by others.
• Upcoming deadlines and collisions.
• A single meaningful creative priority for the day.
• Items the AI can complete without human involvement.

It can also set up recurring tasks (daily, weekly, monthly) so this kind of triage and planning happens automatically.

Visualizing your second brain

For users with a large Obsidian vault or similar note system, ChatGPT can map out your “second brain”: sources, wiki pages, journals, memory pages, and how they all interconnect. It can then generate a visual graph of your ideas and attention patterns, helping you see where your time and focus are actually going.

Custom dashboards for creators and teams

By combining GPT‑5.6 with tools like spreadsheets and dashboards, you can have the assistant build:

• A “today” view that surfaces urgent emails, Slack threads, and tasks.
• An AI intel page that tracks breaking AI news and surfaces relevant items.
• Brand awareness monitors that watch for mentions of your name or brand across the web.
• Traffic and performance overviews for your websites.

These dashboards can be set as your browser’s default start page, giving you a live, AI-curated control center every time you open a tab.

GPT Live: real-time, interruptible voice conversations

OpenAI also launched GPT Live, a new voice mode designed for natural, back-and-forth conversations.

Unlike older voice modes that waited for you to finish speaking, GPT Live can:

• Respond in real time and interrupt when appropriate.
• Be interrupted by you mid-sentence.
• Maintain a more human-like conversational rhythm.

Live translation between languages

One of the most compelling use cases is live translation. You can place a phone with GPT Live running between two people who don’t share a language—for example, English and Hindi—and it will translate both sides in near real time, speaking over each person to keep the conversation flowing.

To enable it in the ChatGPT app, go to Settings → Voice, choose the Live model, and adjust the intelligence slider depending on whether you want faster or smarter responses.

xAI’s Grok 4.5: quietly competitive with top models

xAI released Grok 4.5, a new version of its flagship model. Grok doesn’t get as much attention as OpenAI or Anthropic, but the latest benchmarks show it’s very much in the top tier.

Benchmarks and capabilities

On DeepSWE, the current go-to software engineering benchmark, Grok 4.5 sits just behind Claude Fable Max and GPT‑5.5, and ahead of many other models. On TerminalBench, which measures how well a model can operate a terminal (important for agent use cases), Grok is right up there with GPT‑5.5 High and Fable Max.

These results came out just before the GPT‑5.6 benchmarks, which now take the top spots on both DeepSWE and TerminalBench. Still, Grok 4.5 is clearly competitive for coding and agentic tasks.

Developer-friendly CLI and creative demos

xAI provides a simple command to install a Grok CLI locally, letting you run the model from your terminal without even adding billing details (for now, it can be used for free in that mode).

Grok can also generate interactive visualizations and websites. Given the same “build a website that will impress me” prompt used to test GPT‑5.6, Grok produced a cosmic, starfield-based experience with scale-of-the-universe zooming and constellation drawing. It’s not quite as visually rich as GPT‑5.6’s best efforts, but still impressive.

Pricing for Grok’s API is around $2 per million input tokens and $6 per million output tokens—similar to lower-tier GPT‑5.6 models like Luna and Terra.

Meta’s Muse Spark 1.1 and Muse Image: back in the model race

Meta re-entered the frontier model conversation with Muse Spark 1.1, a capable LLM that, while not state-of-the-art, is competitive with the previous generation of OpenAI and Anthropic models.

Muse Spark performance

On TerminalBench, Muse Spark scores around 80, close to GPT‑5.5 and Opus 4.8. On DeepSWE, it lands in the low 50s, below the newest models but solid enough to be useful for coding and agent workflows.

What’s most striking is the leap from Meta’s earlier outputs. On creative benchmarks like “Buccybench” (an SVG portrait generation test), the jump from older Llama-based models to Muse Spark is dramatic—visually and qualitatively.

Given the prompt to build an impressive website, Muse Spark created a responsive, visually rich page with interactive backgrounds that react to mouse movement, arguably more polished than Grok’s version.

Pricing is around $1.25 per million input tokens and $4.25 per million output tokens, putting it in a similar cost band to Grok and GPT‑5.6 Terra.

Muse Image and Instagram-powered generation

Meta also launched Muse Image, an image generation model integrated into Meta AI. It can create images from text prompts, but the controversial feature is its ability to generate images of real people by @mentioning their Instagram handles.

For example, you can prompt: “Create an image of @username doing X,” and it will synthesize an image of that person in the requested scenario. There are settings to opt out, but rollout is inconsistent and not everyone has the toggle yet, raising serious privacy and consent concerns.

Hyper Agent’s marketplace and B‑roll generator

The team behind Airtable launched Hyper Agent, a platform for building teams of AI agents that can perform complex workflows. This week, they introduced a marketplace where anyone can share reusable “skills” that other users can plug into their own agents.

One example is a B‑roll generator skill:

• You paste in a video script.
• The agent breaks it into shots and maps which visuals should align with which lines.
• It routes prompts to the right video models (e.g., VO for realistic footage, Hyperframes for motion graphics).
• It generates clips, overlays, lower thirds, and even a shot list.
• Finally, it stitches everything together with ffmpeg into a ready-to-use B‑roll reel.

In the Hyper Agent interface, you can watch all assets appear on a canvas, then export the final video or reuse individual clips. The marketplace is built around forking: instead of starting from scratch, you can take agents and skills others have already refined and adapt them to your own workflows.

Anthropic: Fable extension, Co‑work everywhere, and J‑space research

Anthropic had a busy week too, especially around Claude Fable and its Co‑work agent environment.

Fable access extended and limits reset

Claude Fable was originally set to be available only until July 7, but Anthropic extended access to July 12. Many users had already hit their weekly usage limits by the original cutoff, so Anthropic reset those limits—conveniently on the same day GPT‑5.6 launched—giving people more time to experiment with Fable instead of immediately switching over.

Claude Co‑work on mobile and web

Claude Co‑work, Anthropic’s agentic workspace, used to run only on your laptop. If you closed your machine or it went to sleep, the work paused. Now, Co‑work tasks run in the cloud and are accessible from both web and mobile.

That means you can:

• Start a long-running task at your desk.
• Close your laptop.
• Check progress or pick up the output from your phone.

This mirrors the direction OpenAI is taking with the new ChatGPT app, where projects continue running in the background and can be managed across devices.

Claude “Reflect”: your usage, summarized

Anthropic also introduced a “Reflect” feature—essentially a personal analytics view of how you use Claude. It shows:

• Your most active day and peak usage hours.
• Conversation counts over time.
• The main themes or types of work you do with Claude.

It’s a bit like a “Spotify Wrapped” for your AI usage, available in Claude’s web and desktop apps.

J‑space: peeking into an AI’s subconscious

On the research side, Anthropic published work on what they call J‑space. When a model thinks through a problem, you can see its explicit chain of thought (if exposed), but there are also internal activations—concepts and associations it keeps “in mind” without surfacing them directly.

J‑space is Anthropic’s term for this internal representational space. By probing it, they can observe patterns in what the model is implicitly considering, almost like an AI subconscious. It’s early, experimental work, but it opens up fascinating avenues for interpretability and safety.

Google’s video remix and ByteDance’s Seeddream 5.0 Pro

Google and ByteDance also shipped notable media-focused updates.

Google Photos video remix with Gemini

Google added a video remix feature to Google Photos, powered by Gemini Omni. It can:

• Create stylized, shareable video clips from your existing footage.
• Add imaginative effects and transitions.
• Highlight memories in more creative ways than simple slideshows.

This feature is rolling out to users on Google’s AI Plus, Pro, and Ultra plans.

Seeddream 5.0 Pro: ByteDance’s new image model

ByteDance released Seeddream 5.0 Pro, a strong image generation model available through platforms like Leonardo. It stands out for:

• High-quality infographics with dense text and visuals.
• Interactive editing where you can circle regions and specify changes.
• Layer-based outputs similar to Photoshop, letting you rearrange or regenerate individual elements.
• Very photorealistic renders when prompted.

In tests, it produced clean, visually appealing infographics (for example, explaining Anthropic’s J‑space) and convincing portraits, including a recognizable caricature of Gary Busey.

Where things stand now: 5.6, Fable, and the new baseline

The last time the AI community felt a leap this big was the jump from GPT‑3.5 to GPT‑4. For a while after that, progress felt more incremental. With Claude Fable and GPT‑5.6, we’re back to seeing truly massive jumps.

We’re now at the point where you can:

• Hand an AI a detailed product spec and say “build the whole platform.”
• Let it run for a few hours.
• Come back to a working, hosted application that’s 90–98% of what you imagined.

There are still rough edges—UI preferences, small missed requirements—but the fact that a single prompt plus a “keep going until the roadmap is done” command can produce a production-ready system is a huge shift.

In this new landscape, a sensible workflow might look like:

• Use GPT‑5.6 (especially Soul and Terra tiers) as your main builder and day-to-day assistant.
• Use Claude Fable when you want a second opinion on architecture, strategy, or particularly hard reasoning tasks.
• Pull in Grok, Muse Spark, or Seeddream when you want alternative perspectives, cheaper runs, or specific strengths (like visuals or open-source friendliness).

And the wild part: this is the worst these tools will ever be. The pace of improvement is accelerating, and the new ChatGPT super app plus Sites, GPT Live, agent platforms like Hyper Agent, and cross-device workspaces like Claude Co‑work are all early steps toward a world where AI agents routinely build, host, and maintain the software and content we use every day.