July AI News
The New Frontier: Fable 5, GPT 5.6, and Kimi K3
July was the month that three frontier models landed almost on top of each other, and the month the government started showing up in the release process. Here's what actually matters if you're running a business and don't have an hour to spare.
Claude Releases Fable 5 (again)
After export controls, a temporary block, and a genuinely strange few weeks, Anthropic's Fable 5 is live in the Claude model picker. Ask it about cybersecurity or biology, and it routes you down to Opus 4.8. Everything else gets the most capable model available to the public.
Usage Pricing
The free promo ended July 19. Fable 5 now runs on usage credits, and Pro accounts got $100 free to start. It does not auto-bill; you have to opt in through settings to spend more, but expect this tiered pricing model to spread to other model providers.
GPT-5.6 arrived as three models
Sol (flagship, competing with Fable 5), Terra (mid-tier), and Luna (small and cheap), sized like the sun, earth, and moon. Each has an effort slider from extra-high to light, so you dial reasoning depth per query. If that makes OpenAI meaningfully more token-efficient, price might become the next battleground. It was briefly held up by the same new government review that snagged Fable 5, but the real news is in the new features that rolled out with the update (The New ChatGPT Live and ChatGPT Work).
The New ChatGPT Live
Voice used to be three hops (speech to text, then the model, then text to speech), which forced that stilted turn-based rhythm and broke on any background noise. The new version runs in parallel: it listens continuously, drops in "mm-hmm" mid-thought, and says "checking" while researching in the background without stopping the conversation.
Why this matters: AI phone systems, AI Avatars, and customer-facing voice were all held back by that fake feel. That barrier is gone. Only ChatGPT for now, but you can expect the other models to follow soon.
ChatGPT Work: Your new AI Coworker
The Codex app is now the ChatGPT desktop app with a Work mode, a direct answer to Claude Cowork, Copilot Cowork, and Gemini Spark.
The unlock is the desktop, not the model. Point it at a folder and say "this is your working directory." Organize hundreds of files, batch-rename, generate documents, run multi-step tasks against real files. None of that works in a browser tab.
Anthropic reported roughly 90% of what people do in Cowork, and Claude Code isn't coding. It's business operations and content creation. If the code tab looks intimidating, Cowork is the approachable version.
Kimi K3: the largest open model
Moonshot AI (Beijing) released a 2.8-trillion-parameter model and is publishing the weights. It's competitive with the frontier and leads outright on front-end app development, despite China being cut off from the most advanced NVIDIA chips. Constrained on hardware, they innovated on architecture instead. Rumors of a US ban on Chinese models were walked back, but the uncertainty remains.
Honorable mentions
Meta is back with new image and Muse video models, well received after a long quiet stretch.
Microsoft's Frontier Company embeds Microsoft AI engineers inside your organization to build with you.
Grok 4.5 shipped, notable because xAI had been selling compute to Google and Anthropic.
The GPT-6 story, a model was asked whether it could perform a hack. It took a roundabout path. It got the answers. Full coverage next month.
Links to the Stories
Connect with us: Jon Down · Addie Olson · Alec Kauffman
The Summer Automation Series
This summer, we're doing something special called the AI Automation Series, where we'll be going through the Seven Levels of Automation!
Each week, I'll be going through a different level of automation that you can do today to help you get back more time in your day. All summer long, we'll be going through the levels of automation.
Level 1: Software Tools & Data Integrations (7/17/2026)
Level 2: Scripts & Macros (7/24/2026)
Level 3: Workflows (7/31/2026)
Level 4: AI Workflows (8/07/2026)
Level 5: Custom Software (8/14/2026)
Level 6: AI Agents (8/21/2026)
Level 7: Computer Use (8/28/2026)