Matt Wolfe
July 10, 2026
GPT-5.6, ChatGPT Work, and the AI Boom of the Week
TL;DR
OpenAI released GPT-5.6 with near-Fable-level intelligence at lower costs, launched ChatGPT Work as a unified super app, while Meta and xAI surprised with competitive new models—marking a massive leap in AI capabilities across the board.
Key Takeaways
- 1GPT-5.6 represents a major leap over 5.5, delivering near-Fable-level intelligence at significantly lower cost and outperforming on benchmarks like Deep Seek and Terminal Bench
- 2ChatGPT Work unifies Codeex, ChatGPT, and Atlas into one app with dual modes for coding and general productivity, including a personal assistant that can integrate with Gmail, Slack, Google Drive, and other tools
- 3Multiple models launched simultaneously: xAI's Grock 4.5 is competitive on benchmarks, Meta's Llama Spark 1.1 re-enters the race, and Anthropic expanded Claude Co-work to mobile and web
- 4AI agents and no-code tools are maturing rapidly—Hyper Agent's skill marketplace enables sharing of custom agents, while ChatGPT's website builder requires just a single prompt to create fully hosted, shareable sites
- 5The jump in capability feels comparable to GPT-3.5→GPT-4; models can now handle complex full-stack development with minimal iterations and maintain background tasks across devices
- 6Voice mode improvements enable real-time interruption for natural conversation and real-time translation between languages, making AI assistants more conversational and multilingual
- 7Both GPT-5.6 and Fable excel in different ways—Fable stronger on architectural decisions, GPT-5.6 faster at task completion and security; using them in tandem provides complementary benefits
Notable Quotes
“Despite being called GPT 5.6, it feels like a huge leap over GPT 5.5. Like I feel like when we went from 5.3 to 5.4 and 5.4 to 5.5, those were all like really marginal updates. But 5.6, to me, this feels like it should have been GPT 6.”
— Presenter
“The big story around 5.6 is that you're getting like almost as good as Fable Level Intelligence, but for way cheaper.”
— Presenter
“It's kind of blowing my mind what we can do with Fable and GPT 5.6 right now... This is some of the most fun I've had in the last like I don't know year and a half.”
— Presenter
“This is just the worst it's ever going to be. This is just crazy.”
— Presenter
Chapters
1. GPT-5.6 Launch and Capabilities
OpenAI released GPT-5.6, which feels like a generational leap compared to 5.5. The model delivers nearly Fable-level intelligence at much lower costs ($0.77 vs $4 for similar outputs). Benchmarks show it excels at economically valuable tasks, computer use, and software engineering, with Soul Ultra being the most powerful tier.
2. ChatGPT Work: The New Super App
OpenAI unified its products into a single ChatGPT app with Work and Codeex modes. Work mode serves as a personal assistant connecting to Gmail, Slack, Google Drive, and other tools to prioritize tasks and automate workflows. Codeex mode is for development. A new Sites feature allows creating and hosting interactive websites with a single prompt, no database or hosting management required.
3. Personal Assistant Integration and Control Tower
ChatGPT Work can analyze all user activity—emails, calendar, Slack, meeting notes, documents, and project history—to create a daily control tower dashboard. It identifies priorities, upcoming deadlines, actionable items, and can automate recurring tasks. Users can visualize knowledge bases like Obsidian vaults to understand information architecture and attention patterns.
4. Website and Game Generation Demos
GPT-5.6 successfully built complex projects on first try: a full 3D Mega Bunk clone game playable online, and an interactive website with responsive design elements. Meta's Llama Spark and xAI's Grock 4.5 also generated impressive websites and games, though GPT-5.6 generally produced more polished initial outputs requiring fewer iterations.
5. Hyper Agent Marketplace and B-Roll Generation
Hyper Agent launched a skill marketplace allowing users to share and fork custom AI agents. A B-roll generator skill converts video scripts into multi-source video assets, intelligently routing prompts to specialized models (realistic video, motion graphics, charts, text overlays) and assembling them with ffmpeg into complete edited videos.
6. GPT Live Voice and Real-Time Translation
OpenAI's new voice mode enables interruption during conversations, creating more natural dialogue. It excels at real-time translation—placing a phone between two speakers of different languages allows seamless bidirectional translation. Users can choose between quick-response and smart-response modes for different use cases.
7. xAI's Grock 4.5 and Meta's Llama Spark 1.1
xAI surprised with Grock 4.5, achieving competitive scores on Deep Seek (72.7% vs GPT-5.6's 72.7%) and Terminal Bench (83% vs GPT-5.6's 88%). Grock is installable via CLI. Meta re-entered the conversation with Llama Spark 1.1, performing between last-gen (Opus 4.8) and current models. Both are priced competitively near GPT-5.6's Terra tier.
8. Meta's Image Generation and Controversy
Meta released Muse Image, which can generate images of any @mentioned person, even without consent. While free in Meta AI, the feature sparked privacy concerns. Users can disable it in settings, though rollout has been uneven. The model generates quality images but raised ethical questions about identity synthesis without permission.
9. Claude Updates and Anthropic Research
Anthropic extended Fable access and reset weekly limits ahead of GPT-5.6's launch. Claude Co-work now works on mobile and web with cloud-backed task persistence. Anthropic published research on 'JSpace,' exploring AI's latent thought process beyond visible reasoning—essentially mapping the model's 'subconscious' patterns.
10. Closing Thoughts and Future Outlook
The creator emphasizes this is the biggest week in AI in over a year, with jumps comparable to GPT-3.5→GPT-4. GPT-5.6 (daily driver) and Fable (second opinion for complex decisions) are complementary—5.6 executes faster but Fable catches architectural details. Models now handle multi-hour development projects with single prompts, achieving 98% accuracy first try.
Key People & Entities
- OpenAI
- Released GPT-5.6, ChatGPT Work, ChatGPT Live Voice; dominant force in week's announcements
- Anthropic
- Released Claude Fable (extended access), Claude Co-work (mobile/web), published JSpace research
- xAI
- Released Grock 4.5, a competitive open-source-installable model; pricing similar to GPT-5.6 Terra
- Meta
- Released Llama Spark 1.1 model, Muse Image (with privacy concerns), and Google Photos video remix
- Hyper Agent (Airtable team)
- Launched skill marketplace for shared AI agent workflows
- Byte Dance
- Released Seeddream 5.0 Pro image generation model with infographic and layer support
Glossary
- Soul, Terra, Luna (model tiers)
- GPT-5.6 comes in three capability tiers: Soul (most powerful, most expensive), Terra (mid-tier), and Luna (lightweight, most affordable). Users adjust via a slider in ChatGPT.
- Deep Seek (benchmark)
- A benchmark measuring software engineering capability and ability to solve complex coding problems. Used to compare frontier model performance.
- Terminal Bench (benchmark)
- A benchmark evaluating how well models can execute actual terminal/CLI commands on computers, critical for assessing agentic AI systems.
- GDP Val (benchmark)
- A benchmark testing how well models perform on economically valuable tasks in real-world applications.
- Buybench (benchmark)
- A creative/visual benchmark where models generate SVG artwork of faces, with winners ranked on visual quality and aesthetics.
- Claude Co-work
- Anthropic's AI agent feature that executes multi-step development tasks. Now available on web and mobile with cloud backing for background task persistence.
- JSpace (Anthropic research)
- A latent thought space in Claude representing the model's implicit reasoning and patterns—essentially the AI's 'subconscious' thoughts beyond visible reasoning traces.
- Sites (ChatGPT feature)
- A ChatGPT feature allowing users to create, customize, and host interactive websites with a single prompt; handles database, CDN, and hosting automatically.
- Muse Image
- Meta's new image generation model that can generate photorealistic images, infographics, and layer-based editable content; includes ability to generate images of @mentioned people.
- Hyper Agent Marketplace
- A shared skills library for AI agents where users can browse, fork, and customize agent workflows for tasks like B-roll generation, without building from scratch.