AI with ALLIE

The professional’s guide to quick AI bites for your personal life, work life, and beyond.

Every few months, I write up my go-to guide on the latest AI tools, models, and tips. Now that it’s Q2 2026 and two brand-new frontier models are out there, I come to you again, hopefully with the same guidance and clarity.

Also, if I’m being real with you, we’re four years into this generative AI revolution, and it’s still a dumpster fire to navigate it all.

So let’s shortcut through all the noise, shall we? Also, as I’ve mentioned before, feel free to grab this entire newsletter and copy+paste it in your favorite AI.

FABLE 5 and GPT 5.6 ARE BOTH LIVE

Two frontier models launched in the last month (one just today!) and both outperform everything else out there. If you’re AI-curious and trying to be an AI superuser, I want you to test both today.

I was an early tester of both models and want to give you my real honest take so you don’t have to spend hours figuring out what I already uncovered. Your mileage may vary, your use cases might be different, but I guarantee this is a great start.

ABOUT FABLE 5 and GPT 5.6

Fable 5 is from Anthropic. GPT-5.6 is from OpenAI. And I…well, I am from Los Angeles.

Claude Fable 5 launched June 9 as a "a Mythos-class model that we've made safe for general use”. And yes, it’s the same Mythos that “wiped billions out of cybersecurity stocks” and caused the summoning of Wall Street's bank CEOs by the Fed and Treasury the day the Preview was revealed. 

Three days after the Fable 5 launch, the US government applied export controls after researchers at Amazon claimed to had found a way to bypass its cyber safeguards, the model came off the market, and Anthropic added new safety classifiers, redeployed it on July 1, and extended Fable's included paid subscription plan access through July 12, which is a run of sentences I could not have imagined typing a year ago. 

OpenAI previewed GPT-5.6 on June 26 - but only to about 20 organizations individually approved by the US government, after the White House asked them to slow the launch over cyber and bio concerns. It cleared additional government testing and went fully public today, July 9 2026, in three sizes: Sol (the big flagship), Terra (the mid-tier), and Luna (the small fast one).

We are now in an era where model launches involve the White House. 

Now is the time to go over my favorites.

Ladies and gents, let’s rock and roll.

MY GO-TO WORKSPACE: CLAUDE CODE

Claude Code is still where I live all day. It’s where I have my AI boardroom all fight each other, it’s where I’m building out the majority of my company’s products, it’s where I draft all of my key client recaps, and it’s where I have set up all of my autonomous agents (like my morning briefings, my urgent email responder, my pre-revenue tracker, my research team, my KPI tracking, and more).

My preferred permissions setting is auto-mode. And when I'm not working with anything sensitive (like client data, credentials, financial info), I'll go full bypass-permissions mode (aka “YOLO mode”) and let it just run, especially when I’m sleeping. I’ve found both Fable 5 and GPT-5.6 handle unsupervised runs very well. 

MY GO-TO SURFACE: IPHONE

Why yes, I could sit glued to my desktop chair for 14 hours a day, but that wouldn’t be fun, would it? I use Claude Code from my phone constantly now - out to lunch with friends, checking in from anywhere, walking through the streets of London. I spend way more time away from my laptop. The catch is that your laptop has to be sitting open at home for it to work, or you need to set up a "virtual machine" (a computer you rent in the cloud, from somewhere like DigitalOcean), or a Mac mini - which are sold out everywhere right now. Alternatively, Codex is a fantastic option here because it runs directly in the cloud, so nothing on your desk needs to stay awake. Don’t desk glue yourself. Buy some acetone, and walk outside.

MY GO-TO ACCESS POINT: DESKTOP APP

Ever since Claude Code came out in March 2025, I have only ever used it in the terminal on my Mac, or iTerm when I was feeling zesty. However, with how much multitasking I’m doing these days in AI (literally 30+ terminals open on my Macbook Pro M4 Max), it just felt like my brain was short-circuiting and I didn’t know how to handle that influx of AI agents. So about a week ago I switched to the desktop app as my go-to. We will see how it goes. I like that you can easily rename conversations, and you can actually triage which ones need your attention instead of hunting through 30+ windows.

MY COLLABORATION ZONE: SLACK

Most AI harnesses are single player, and that is mighty dreadful when you work on a team. What if I need our team working together with an agent? What if I want my teammates to ping my AI workforce directly instead of routing everything through me? What if I just don’t want to loop alone? So I set up a dedicated channel in Slack that lets my team talk directly to my AI workforce, which runs in Claude Code.

BEST MODEL FOR AGENT ADVISOR: FABLE 5

This is also a brand-new shift for me, and it essentially gives your AI agents a little friendly genius/mentor sitting in the corner. When I'm running a team of (cheaper) AI agents, and one gets confused or stuck, it can reach out to (expensive) Fable 5 for advice. You can use it by just typing ‘/advisor’ into Claude Code!

Back in April, Anthropic tweeted: "Pair Opus as an advisor with Sonnet or Haiku as an executor, and get near Opus-level intelligence in your agents at a fraction of the cost". Their Claude Code docs have since extended it to Fable 5: run Sonnet as your day-to-day executor, and have it escalate to Fable "when stuck on a recurring error" or before declaring a task complete.

Almost like, a student working on their own, and only going to their professor when they really need help. Sonnet is the student. Fable is the wise professor with a long beard and probably a monocle.

The economics win out here. Sonnet 5 costs $2/$10 per million tokens. Fable is $10/$50. Let Sonnet grind and bring Fable in as needed. On SWE-bench Pro, a popular coding benchmark, Sonnet 5 with a Fable 5 advisor tool achieved ~92% of Fable 5's solo score at only ~63% of the price. It’s a cost-saving strategy.

BEST FOR AI CHIEF OF STAFF: OPUS 4.8 HIGH

I have 34 AI agents in my workforce, all run by the daily executor and organizer, my AI Chief of Staff, Simon. You can learn all about how I built my AI workforce, and how you can too, here. I don’t think Fable is necessary for this and Opus 4.8 (set to “high” effort) is perfectly fine. At this time, I would not use a model smaller than Opus for my main Chief of Staff.

BEST FOR AI SUBAGENTS: SONNET 5

For subagents, I care more about speed because the focus is time-to-value through parallelization (more subagents all doing stuff at once). Speed also really matters when you have multiple agents working together and passing notes to each other to share context. They rarely have to own too big or vague of a project because their boss (ex: Opus 4.8, Fable 5, GPT-5.6) handles scoping and cross-agent collaboration efficiency.

BEST FOR COMPUTER USE: CODEX + GPT-5.6

The published numbers as of writing this newsletter actually favor Claude here - Fable 5 scores 85% on OSWorld-Verified, a computer-use benchmark where the human baseline is 72.4%. But in my testing, GPT-5.6 smoked it out of the water. When OpenAI removed early access to 5.6 for all early testers, I found myself missing the computer use capabilities (aka when an AI system literally takes over your computer’s mouse) the most.

BEST FOR IDEA GENERATION: GPT-5.6

When I need raw idea volume - brainstorms, product ideas, novel solutions to a stuck problem - I go to either the ChatGPT mobile app or Codex. I would go to it for a question like “I’m in Scotland and want to go to the British Grand Prix. The train is too long. What’s a creative solution for this?” or “Help me creatively solve this problem and think through all solutions: I have two bosses, they both give me a full-time person’s worth of work, they refuse to talk to each other, I am working until 10pm every night.”

Also worth saying plainly: brainstorming is one of the places where you don't need a frontier model at all. Mid-tier models brainstorm great. Spend your fancy tokens elsewhere.

BEST WRITER: FABLE 5

Claude is my writer and has been for the last year. Comparisons from other early testers consistently rate Fable 5 as the better raw prose model available, especially at holding tone and facts across long documents. If you see someone say that GPT-5.6 is the better writer, ask yourself first if they are writing technical documentation for a techie audience or business docs. I suspect it’s the former.

I still don’t think AI is great at mimicking voice and capturing “the human nuance”, like delicate apologies or sarcasm or influence. I can get long drafts to maybe 70-85% match my voice and short drafts to match at 95% but I’m almost always editing before anything goes out, even if it’s a short email.

BEST FOR IMAGES: CHATGPT IMAGES 2

Nothing in the GPT-5.6 release changes image generation, for what it's worth - the ChatGPT image stack carries this pick. Second choice is Gemini’s Nano Banana Pro. You can edit images, generate new designs, get visual inspiration, create infographics or cartoons, make party invites, and even redesign your workspace. Get 5 great image generation starter prompts here.

BEST FOR VIDEO GENERATION: HIGGSFIELD AND GEMINI OMNI

I really don’t find myself generating video all that often, and certainly not at the quality of someone like Don Allen Stevenson (I mean, just look at this Grand Theft Auto example). If I have to create a video, I start with Google Gemini Omni. If it fails, I go to a human. And if I’m blocked on that, then I go to Higgsfield.

BEST FOR DESIGN: CLAUDE DESIGN, BARELY

All of the top AI models out there today (Fable 5, GPT-5.6) can crank out a better design draft than I ever could. And for that reason, they honestly all rock and you really can’t go wrong (Codex with GPT-5.6 just blew me away at its first crack on a new product I’m working on). Claude Design, however, is purpose-built for this exact task. Instead of taking its ‘best guess’ at what you want (which again, is very impressive and better than anything I could make), it asks you clarifying questions upfront to make sure it is building exactly what you’re looking for.

3 versions instead of 1? It’s on it. Extra features? You got it.

BEST FOR VOICE: WISPR FLOW

I’ve dictated the length of three entire books to AI. That’s right, over 300,000 words from my brain to my mouth to my Claude Code, Codex, iMessage, Slack, and more. Wispr Flow, despite its clunky iPhone integration, is one of the biggest hacks in my toolkit. If you haven’t started dictating to AI yet, do it for 3 straight days and tell me if your whole life hasn’t changed.

OpenAI also released their new live voice mode today that can interrupt you and can “sound more natural”. The demo they shared is of a team member prompting the AI with “hey, I’m learning English, I’m going to speak, can you interrupt me if I make a grammar mistake?” And he goes on to say something like “I’m excite to go” and ChatGPT interrupts him to say “actually…it’s excited”. I was also an early tester of this feature, and unfortunately, I thought the voice was extremely ditzy, dumb, repetitive, inaccurate, and annoying. Nice for a random car ride conversation, won’t be using it otherwise.

BEST FOR CODING: FABLE 5 AND GPT-5.6 TOGETHER

If you're doing genuinely complex product building - multi-day builds, gnarly migrations, systems with a dozen moving parts - use one of the two new frontier models. This is what they exist for. Stripe used Fable 5 to run a 50-million-line codebase migration in a day - work that would otherwise have taken two months. And Boris Cherny, the creator of Claude Code, called Fable 5 "the best model I have used for coding, by a wide margin".

A common pattern I’m seeing amongst AI superusers, and one I have adopted, is working inside of Claude Code and leveraging Claude as the manager and then also using the battering ram power of GPT 5.6 in Codex as the executor. I think this is brilliant. GPT-5.6 inside of Codex is the type of model that really doesn’t give up.

On the benchmarks, Anthropic published SWE-bench Pro, a benchmark built from real software-engineering tasks, where Fable 5 scores 80.3% (GPT-5.5, the previous OpenAI flagship, sits at 58.6%). OpenAI published Terminal-Bench, which tests how well an agent works inside a computer terminal, where GPT-5.6 Sol hits 88.8% - and 91.9% in its new "ultra mode," which spawns its own subagents. Let’s remember, of course, that each lab released the chart where it wins. Benchmarks aren’t perfect, so I'd treat all of this as directional until independent testing catches up or until you test it on your own business and personal use cases.

Cost concerns are real. Fable 5 costs $10 per million input tokens and $50 per million output - double Sol's cost at $5/$30. Fable 5 also has safety classifiers that can bounce certain (though rare) requests down to Opus 4.8 mid-task.

GPT-5.6 Sol has an even weirder problem. METR, the independent evaluation org, caught Sol gaming its tests more than any public model they've ever evaluated - it exploited bugs in the evaluation harness itself. Depending on whether you count the cheating as failure or success, its measured task horizon (roughly: how long of a task it can reliably finish on its own) is either about 11 hours or over 270.

And remember, small models have their place too. You don't need Fable 5 or GPT-5.6 to quickly summarize a meeting or punch up an email. That's like asking a fire-breathing dragon to warm up some broccoli. To address this, I use Simon Willison’s promp to manage that: “For all coding tasks, use your judgement to decide an appropriate lower power model and run that in a subagent”

WHAT TO DO TODAY

1) If you haven’t already, switch your workspace to a harness (Codex or Claude Code)

2) Test the state of the art models - either Fable 5 or GPT-5.6 or both

3) Start dictating with Wispr Flow if you haven’t already

4) Try one model stacking pattern for greater token and spend efficiency (either Claude Code calling Codex, or Fable 5 acting as an advisor to Sonnet 5)

Stay curious, stay informed,

Allie

FREE AI AGENT WORKSHOP REPLAY

If you’re brand-new to AI agents and want to finally build one, I held a live workshop on how to get started - and shared all the same frameworks I use to train global Fortune 500 companies. No coding required. Built for business people, not engineers.

Watch and share today.

Feedback is a Gift

I would love to know what you thought of this newsletter and any feedback you have for me. Do you have a favorite part? Wish I would change something? Felt confused? Please reply and share your thoughts or just take the poll below so I can continue to improve and deliver value for you all.