In this issue
Nine days, three frontier labs, three new model families. OpenAI shipped GPT-5.6 this week along with ChatGPT Work, an agent that takes a goal, breaks it into steps, and stays on the job for hours across your apps and files. SpaceXAI answered with Grok 4.5, Meta showed up with Muse Spark 1.1, and suddenly frontier-class AI costs anywhere from $1 to $50 per million tokens depending on where you shop.
The top models are now separated by a single benchmark point and a 3x price gap, which makes picking models per task a real operator skill. And the demo of the week shows how far the ceiling moved. One prompt, one Mac, and GPT-5.6 trained a working language model from scratch. Let us get into it.
Topics of the day:
OpenAI ships GPT-5.6 and the ChatGPT Work agent
GPT-5.6 trains a language model from one prompt
Grok 4.5 lands at $2 per million tokens
Curated reads on local models and AI-native hiring
Meta's Muse Spark 1.1 goes for cheap and everywhere
The new scoreboard vs Claude Fable 5
The Shortlist: GPT-Live voice, Copilot upgrade, Claude reflection
OpenAI ships GPT-5.6 and turns ChatGPT into a workday agent
What's happening: OpenAI released the GPT-5.6 family this week, flagship Sol plus mid-tier Terra and budget Luna, and paired it with ChatGPT Work, an agent that gathers information across your apps and files, breaks a project into steps, and stays with it for hours.
In practice:
Give ChatGPT Work a task you already know well, like a month-end budget variance or a campaign brief, so you can judge its output honestly.
Scheduled Tasks keep projects moving while you are away, turning new Slack and Teams messages into updated docs and slides.
Pricing dropped hard: Sol runs $5 per million input tokens, and OpenAI says the $1 Luna tier beats Claude Opus 4.8 on coding benchmarks.
Bottom line: ChatGPT stopped being a chat window this week. Hand it one recurring workflow and check the work like you would a new hire's.
You can now vibe code your own language model
What's happening: Pietro Schirano, the MagicPath CEO who used to design at Anthropic, gave GPT-5.6 a single prompt and it built the entire training pipeline and trained a small language model from scratch on his iMessage history, locally on his Mac. It now drafts replies in his writing style.
In practice:
The demo collapses what used to be a data science project into an afternoon, no ML team, no cloud GPUs, no vendor.
His messages never left his laptop, a pattern worth copying for anything you would not paste into a chatbot.
Try it on a corpus you own, like sent emails or old proposals, and you get a private model that drafts in your voice.
Bottom line: I always thought custom models would stay an enterprise purchase. Run this experiment on your own sent folder before you pay anyone for a fine-tuned brand voice.
SpaceXAI's Grok 4.5 starts a price war
What's happening: SpaceXAI launched Grok 4.5, the first model it trained together with Cursor after the $60 billion acquisition, benchmarking near Claude Opus 4.8 and GPT-5.5 while priced at $2 and $6 per million tokens and running at 80 tokens per second.
In practice:
Usage is temporarily free inside Cursor and Grok Build, so your dev team can trial it this week without touching the budget.
Re-quote bulk jobs like tagging, summarization, and data cleanup at these prices, the same work costs a fraction of Opus rates.
EU access does not arrive until mid-July, so European teams should line up the eval now and flip the switch when it lands.
Bottom line: Grok was a punchline not long ago. Cheap, fast, and good enough wins a lot of everyday work, and Musk says a bigger model lands next month.
Read Later
Running local models is good now - Vicki Boykis argues local models finally earn their keep, at roughly 75% of frontier accuracy on coding, a neat companion to today's iMessage story.
OpenHuman - An open-source assistant with persistent local memory and multi-model orchestration, built to be an evolving agent instead of a stateless chatbot.
Talk to a friend's Codex - A short demo of chatting directly with someone else's Codex agent, worth 30 seconds of your scroll.
The Notion Sales Agent Build Challenge - Notion is recruiting sales development reps through an agent build challenge, a glimpse of what AI-native hiring looks like now.
Meta's Muse Spark 1.1 goes for cheap and everywhere
What's happening: Meta released Muse Spark 1.1, the first flagship update from Alexandr Wang's Superintelligence Labs, claiming wins over Google's latest Gemini on coding and reasoning, with Zuckerberg promising "strong agentic and multimodal models at very low cost."
In practice:
Meta concedes it still trails GPT-5.6 and Claude Fable 5 on coding, so slot Muse into the budget tier of your stack, not the front line.
The companion Muse Image model already sits inside Instagram and WhatsApp, putting creative generation directly in the channels your campaigns run on.
Muse Spark 1.1 is in public preview with aggressive pricing promised, so the cost of good-enough AI keeps falling either way.
Bottom line: Meta is not chasing the benchmark crown, it is putting decent models everywhere its apps already live. For most marketing teams that matters more than a leaderboard.
The new scoreboard: everyone prices against Claude Fable 5
What's happening: Independent tester Artificial Analysis puts GPT-5.6 Sol within one point of Claude Fable 5 on its aggregated Intelligence Index at about a third of the cost per task, while Fable 5 keeps the top spot overall and on the hardest office-work benchmark.
In practice:
Route by job: Fable 5 where sustained judgment carries real money, Sol for agentic coding, Grok 4.5 or Luna for high-volume cheap work.
Frontier-class pricing now spans $1 to $50 per million tokens, which makes model routing a genuine budget lever instead of a nerd preference.
I built Model Fit for exactly this, benchmark the contenders on your own work before you switch anything.
Bottom line: The gap at the top is one point, the gap in price is 3x. Pick per task, not per brand.
The Shortlist
OpenAI released GPT-Live-1, full-duplex voice models that listen while they speak, handle interruptions naturally, and enable live translation, worth testing on support and sales calls.
Microsoft made GPT-5.6 the preferred model in Microsoft 365 Copilot, so Word, Excel, and PowerPoint quietly picked up this week's frontier upgrade for anyone already paying for it.
Anthropic launched a reflection feature that tracks and visualizes how you use Claude, so you can decide whether that time actually matches your goals.
This newsletter is where I (Kwadwo) share products, articles, and links that I find useful and interesting, mostly around AI. I focus on tools and solutions that bring real value to people in everyday jobs, not just tech insiders.
