The ‘Best of the Rest’ AI Models. ARD #116

Michael Parekh AI: Reset to Zero

15 hours ago

18 MIN READ

Today’s theme: the ‘best of the rest’ AI models. This week the two top US frontier models came off ice — Anthropic’s Fable 5, freed after the three-week ‘Blip 2.0’ stoppage, and OpenAI’s GPT-5.6, freed from its gated release. Now we’re seeing the best of the rest: new models from Meta and SpaceXAI, plus notable extras from Anthropic and OpenAI to make their top products easier and better to use. Three events for the AI Tech Wave — each with my Take first, then my Overall Take.

(1) Meta Releases Muse Spark 1.1 — the First Fruits of Its Multi-Billion-Dollar Meta Superintelligence Labs (MSL) Effort

MP TAKE: This is the first major iteration of Meta’s AI models since its big switch last year — tens of billions of dollars poured into AI talent, starting with the ~$14 billion Scale AI acquihire that brought in Alexandr Wang. Mark Zuckerberg went pedal-to-the-metal hiring the best AI researchers he could get from anywhere — OpenAI, Apple, everywhere — and, notably, over a third of them are Chinese-born and/or Chinese-American. The result is Muse Spark (they’d already shipped 1.0 a few weeks ago; this is 1.1), now roughly a 1.5-trillion-parameter model that goes more head-to-head with the current — not the latest — top models from Anthropic and OpenAI on most baseline benchmarks. It does a few things better, especially on agents (the race everyone’s focused on) and images.

And Meta being Meta, they’re using Muse Spark to augment their advertising tools — the way their bread is buttered — letting advertisers create more dramatic AI-generated visuals and pitches to monetize the 3.5 billion people on Instagram, Facebook and WhatsApp. The bigger tell is the business model. Mark talked about making Muse Spark 1.1 available via APIs to developers and businesses at price points a quarter or less of the state-of-the-art models from the two major competitors. That’s the first real inroad on what I’ve been calling — tongue in cheek — ‘ZWS,’ Zuck Web Services: Meta’s AWS-style move to rent out its excess AI data-center capacity (each of these players is spending $145+ billion-plus a year building it out) until it needs that capacity for its own use. He’s not far behind Elon Musk’s ‘EWS,’ Elon Web Services, at SpaceX. And the ready customers are Anthropic and OpenAI themselves, who need emergency compute on their rocket ride to their mega-AI IPOs — Anthropic is already at a $60 billion ARR this month, OpenAI not far behind. It’s a notable pivot for a company that built its reputation on open-source Llama — Muse Spark 1.1 is closed and pay-to-use at scale.

Sources, in narrative order: Axios — Meta updates its Spark model, releases developer version. Bloomberg — Zuckerberg pledges ‘aggressive’ pricing with Meta’s first pay-to-use AI. The Verge — Meta says its new AI model is ready to compete on coding. For longtime readers, in narrative order: today’s companion AI-RTZ #1142; ‘Meta’s New AI Org coming into focus’ in AI-RTZ #982; ‘Meta & Zuck aiming for the AI Clouds’ in AI-RTZ #1135; ‘Surveying mid-year AI options for the Mag 7 and beyond’ in AI-RTZ #769; and yesterday’s ‘AI Price Umbrellas Getting Larger’ in ARD #114.

(2) SpaceXAI Releases Grok 4.5 — with Technologies from the $60 Billion Cursor Acquisition

MP TAKE: SpaceX is not far behind. They released their ‘best-of-the-rest’, Grok 4.5, with technologies from their $60 billion acquisition of AI-coding company Cursor — a transaction that has yet to close (expected in the third quarter), though the two have been working together. Grok 4.5 is supposed to be a lot better not just at coding but at specific domains like legal and finance. There’s a real technical story in the weeds here: the training data behind Grok 4.5 is the coding data Cursor accumulated over its last three years, with millions of developers using their AI Coding tools— much of it riding on top of Anthropic’s Claude Code. Now that the two companies are more separated, Cursor actually uses some Chinese models for its core functions These models are also priced very attractively relative to the top models, available via APIs and à la carte pricing, just like Fable 5 and GPT-5.6.

At a high level, from Elon’s point of view, this is SpaceXAI’s best foot forward yet to move closer to the two top dogs — it took a ~$1.75 trillion mega-AI IPO in June and the Cursor deal to get here. Not to mention the relentless building of gigawatt scale AI data centers. But Elon has a lot of wood to chop relative to his immediate peers in Meta and Google, and a lot more work to get really close to OpenAI and Anthropic in coding and enterprise. Grok 4.5 also emphasizes cybersecurity — an area of focus for the US government of late. And watch the Cursor angle: The Information reports Cursor is building its own agentic product to go head-to-head with Anthropic’s Claude Cowork. Musk is not sitting on his laurels with 4.5 — there’s a lot more to come.

Sources, in narrative order: Axios — Musk’s SpaceXAI releases new model, Grok 4.5. Bloomberg — SpaceXAI, Cursor unveil Grok AI model for coding, legal and finance tasks. The Information — Cursor is developing an AI agent to compete with Claude Cowork. For longtime readers: ‘SpaceX/xAI IPO filing outlines Elon’s boundless AI ambitions’ in AI-RTZ #1066.

(3) Anthropic & OpenAI Roll Out AI Extras Alongside Fable 5 and GPT-5.6

MP TAKE: Both are pedal-to-the-metal now that they have the ‘Go’ green light from the US government on the security and gating front — and neither is sitting on its laurels. On Anthropic’s side, they released ‘Reflection,’ a tool that gives people who use their top products — Claude Code and Cowork (which I use a lot) — a dashboard on everything you’ve been doing over the last few days, to help you organize and plan your use of the tools. On OpenAI’s side, there’s ‘ChatGPT Live’ — a new full-duplex voice capability where you can talk over each other and the system still hears you and reacts to changes. There are a couple of charming commercials for it featuring senior ladies having fun with the voice interaction — they remind me of the ‘Where’s the Beef?’ ads from the burger companies decades ago. Worth a watch. OpenAI also shipped a ‘ChatGPT Work’ agent.

The one big other move: OpenAI is taking a step back from AI browsers. This was a whole category I wrote about a lot last year — Perplexity and others racing to build browsers to compete with Chrome and Safari. But OpenAI has now fused its Codex coding tool with ChatGPT into one app, building super-app capabilities that work on top of any browser you use — so they don’t need their own browser, Atlas, anymore. It confirms a thesis I’ve held for a while: where we’re going, we won’t need the browser as we’ve known it for 25 years — not because a rival browser wins, but because chat, agents and voice become the interface, and these systems increasingly understand how to use your computer better than you do.

Sources, in narrative order: Axios — Anthropic’s Reflection AI gets its screen-time moment; Axios — OpenAI releases GPT-5.6 and ChatGPT Work tool. OpenAI — GPT-5.6: frontier intelligence that scales and the all-new ChatGPT Voice. 9to5Mac — OpenAI unveils ChatGPT Work agent; GPT-5.6 models now available and OpenAI is discontinuing ChatGPT Atlas, its standalone desktop browser. For longtime readers: ‘AI browsers are about a whole new way to use the web’ in AI-RTZ #733.

MP OVERALL TAKE

These are early days — but exciting ones. The early reviews on the top two models have been raving: developers love Fable 5 and GPT-5.6. Of course, some a bit more critical on the OpenAI move to fuse their Codex and ChatGPT apps, both as standalone apps and on the web. Doing their best to compete with the equivalent runaway successes, Anthropic’s Claude Code and Cowork. Confusing, I know. But these are early days for these UI/UX developers at both companies.

There’s a lot of novelty factor enthusiasm and praise for most who are using the best of the best AI models from both Anthropic and OpenAI— tons of YouTube videos of developers building clones of Excel and Minecraft, with the models running multiple AI Agents for days on a single task. Burning gobs of AI compute dollars. Things you and I wouldn’t normally do, at a cost of huge amounts of compute. These early tests typically run $50,000, $100,000 or more in compute cost — imagine a weekly bill like that to really use the top-top-top models. This is not for regular people, and even companies are shrinking from that level of ‘token maxxing’ or ‘token budgeting.’ Even Tesla’s AI-compute budget is now $200 a week per employee. So about $10,000 a year per head. It all has to get a lot more efficient.

On the ranking: GPT-5.6 is generally seen as the top of the current generation. Anthropic’s Fable 5 — rumored at five to ten trillion parameters, three times or more the size — is in a league by itself, the first of the next generation. The best of the rest from Meta and SpaceX are at roughly 1.5-trillion parameters (no one has released the exact numbers or weights), and they run on data centers, not local devices, because of their size. And there’s a huge ‘my God, the parrot can talk — or the horse can count’ moment here: everyone is startled by how far these things have come. For that reason we should take a breath and appreciate the technical results of these gobs of investment. But we’re using total brute force, and all of it has to get far more efficient — at the mainframe/data-center level first, and eventually down to local devices like your phone. It’ll be a long time before we run data-center-class models on local devices — years, especially with RAMageddon and the cost of memory. This is a crawl-crawl-walk-walk-run thing, over five, ten, fifteen years — not overnight, regardless of what new companies selling new gadgets tell you.

For regular people, what you have is good enough — Google Gemini is good enough; Apple Siri AI, coming in September, will be more than good enough on local data and ‘un-metered’ AI agentic tasks, with lots of privacy. Fable 5 and its peers are really designed for the top enterprise and developer customers, not for us — yet. What’s encouraging is the competition, and not just from the US best-of-the-rest. The ones I haven’t covered are the host of open-source AI companies in China — everyone from Zhipu (ZPU) to MiniMax, Qwen 3.7 from Alibaba, and many others — with trillion-plus-parameter open-source models that a lot of US companies are also using (assuming US politicians don’t ban or curb that, which I talked about yesterday). Add Nvidia’s Nemotron and whatever Apple does on the consumer side, and you can see why this is interesting. And barely beginning. We had the government’s ‘Blip 2.0’ slow things down for three weeks, but we are off to the races again — with the best of the rest, and the best from the best.

Gadget AI — Samsung Readies Its Wide Foldable + Wearables for July 22

MP Take: I talked a couple of days ago about the Google Pixel lineup refreshing in August, with Apple’s iPhone 18 series coming in September under new CEO John Ternus. In the meantime, Samsung — the biggest maker and seller of Android phones in the US, and the US leader in foldables — keeps pressing ahead. Its Fold 7 from last year is remarkably thin; it’s arguably the best foldable in the market, and it turns into a mini-tablet. But these are expensive — with today’s RAM prices, $2,000 or more. Samsung briefly sold a three-screen version north of $3,000+, then pulled it because they couldn’t make money given rising component costs. Now Apple is finally rumored to join the foldable race — likely with a wider, ‘passport-size’ format (as opposed to a candy-bar phone). That wide format is one I’ve loved since the first Google Pixel Fold: shorter, but wider, with more readable left-to-right content. So Samsung wants to get ahead of Apple with its own wide foldable, rolling out July 22 alongside new wearables — the Galaxy Z Fold 8 and Z Flip 8, and the Galaxy Watch 9 and Watch 2 Ultra.

It’s very geeky — for early adopters — and the prices will be nosebleed, $2,000-plus. But these matter because of the software innovation that has to happen when Android (and hopefully iOS) software runs across two screens instead of one. Samsung pioneered foldables in 2019 and holds roughly two-thirds of the market. Apple’s first fold is rumored at a million units or less — a showcase, like the Vision Pro was a few years ago at $3,500, really for early adopters. Right now every one of these companies, Apple included, is in the showcase stage. For gadget freaks, this is Christmas in the middle of the year — but still a nice-to-have, not a must-have.

Sources, in narrative order: The Verge — Samsung will launch its new wide foldable on July 22nd. Engadget — Official renders for Samsung’s Galaxy Z Flip 8, Watch 9, Watch 2 Ultra and Z Fold 8. For longtime readers: ‘Samsung vs Apple global AI smartphone race’ in AI-RTZ #958.

Questions

Q1 — What’s Michael’s primary appeal in foldable phones?

It’s nice to have a phone that, in a pinch, turns into a tablet — so you can watch YouTube and other video on a wider screen, or do more productivity work across two screens where the software allows. That mini-tablet form factor is useful. But recognize that most of the time you just use the outer screen — roughly 80% — and you’re more than happy, particularly now that these have gotten so thin. The Samsung foldable is basically as thin as — not even a little thinner than — my iPhone 17 Pro Max. So it’s a nice-to-have, not a must-have for now, especially for the price, and given the limited software UI/UX innovation to truly take advantage of the bigger screen. A cool novelty item for early adopters.

Q2 — What’s Michael’s biggest criticism of the foldables?

The one thing people fixate on is the crease. It’s not that noticeable once you start using the phone, but it’s the sore point. The latest innovations in Asia — where three or four Chinese companies have their own models, in many cases with better technology than ours — are taking the crease and making it disappear through novel display tech that can fold. Those aren’t yet available in US markets. My other big criticism is that there isn’t enough innovation on the OS — the way apps take advantage of two screens versus one. The fact that Apple is entering the foldable space means they think they can do creative things to make it more mainstream — hopefully with interesting iOS work on that front.

Wrapping up

Today’s AI-RTZ #1143 — Alibaba as the open-vs-closed AI ‘canary in the coal mine’ — is on Alibaba, whose Qwen models are considered among the best open-source models out there, ahead of even DeepSeek and others. But like Meta, Alibaba is transitioning toward closed-source to try to monetize more around its technology. It’s an interesting case study with real implications for the global AI market. Recommended as today’s reading.

This weekend — the Saturday Weekly AI Roundup, and Sunday, The Bigger Picture. Have a great weekend.

Thanks for joining us today, AI Curious Folk. Stay tuned.

— MP

Full Source Reading —

For the broader context, see the canonical sources for ARD 116 — in today’s narrative order:

Clips from today

Clip 1 — Apple’s Foldable Phone: A High-End Novelty

Watch on YouTube Shorts

Samsung launches a wide foldable plus new watches July 22, getting ahead of Apple’s first fold — expected this year, likely a low-volume, Vision-Pro-style showcase.

MP Take: I’ve been a long-time foldable adopter — intriguing but expensive, more so under RAMageddon. People use the outer screen 80% of the time; the wide screen shines only occasionally. A cool novelty for early adopters, not yet must-have.

Clip 2 — AI’s ‘Parrot Can Talk’ Moment

Watch on YouTube Shorts

The latest AI models spark real wonder — developers cloning Excel or Minecraft over a weekend — but the top-tier runs cost $50K-100K a week in compute.

MP Take: It’s a ‘my God, the parrot can talk’ moment — worth taking a breath to appreciate. But all of it has to get far more efficient, from data centers down to local devices. A crawl-walk-run over 5-10-15 years, not overnight.

Clip 3 — Meta’s Muse Spark 1.1 vs the Top AI Models

Watch on YouTube Shorts

Meta releases Muse Spark 1.1 — the first real fruit of its multi-billion-dollar Superintelligence Labs effort under Alexandr Wang. A ~1.5T-parameter model, stronger on agents, now offered via API.

MP Take: The tell is pricing — Meta’s first serious pay-to-use API at ~25% of the top models’ cost, the monetization arm of its AWS-style cloud, ‘ZWS’ (Zuck Web Services). A notable pivot for a company built on open-source Llama.

Clip 4 — AI Browsers: OpenAI Steps Back

Watch on YouTube Shorts

OpenAI is discontinuing its Atlas AI browser, fusing Codex into a ChatGPT super-app that runs on top of any browser — Chrome or Safari.

MP Take: This confirms a thesis I’ve held — where we’re going, we won’t need the browser as we’ve known it for 25 years. Not because a rival browser wins, but because chat, agents and voice become the interface instead.

Clip 5 — AI Models: Early Days, High Costs

Watch on YouTube Shorts

Running the top models at full tilt can cost $50K-100K a week in compute — so much that even Tesla now caps AI spend at $200 a month per employee.

MP Take: The wonder is real, but this is brute force. It all has to get far more efficient before it reaches regular people. For now, what you have is good enough — Gemini’s good enough, and Siri AI this fall will be more than good enough.

About AI Ramblings Daily (ARD), and AI-RTZ

Both are daily. Both are free. Both are about AI. But they’re different mediums carrying different messages.

AI-RTZ is the morning text — a deeper written take on one idea, published by at least 5 AM EST. Today: post #1143.

AI Ramblings Daily is the afternoon video + podcast — my ad hoc takes and perspective on the day’s AI issues & news flow, around 20 minutes, with short 1-2 minute clips for quick topic views. Today: episode #116.

Subscribe to either or both on michaelparekh.substack.com. They run as separate Sections you can opt into or out of.

Links used in today’s show (already embedded inline above; listed here for reference)

Take 1 — Meta’s Muse Spark 1.1 (MSL):

Take 2 — SpaceXAI’s Grok 4.5 (+ Cursor):

Take 3 — Anthropic & OpenAI’s Extras (Fable 5 / GPT-5.6):

Gadget AI — Samsung July 22 Wide Foldable + Wearables:

Companion text:

AI-RTZ #1143 — Alibaba open vs closed AI ‘canary in the coalmine’

(NOTE: The discussions here are for information purposes only, and not meant as investment advice at any time. Thanks for joining us here.)

Subscribe to AI: Reset to Zero for daily AI Ramblings + Sunday Bigger Picture posts.

The ‘Best of the Rest’ AI Models. ARD #116

Share

(1) Meta Releases Muse Spark 1.1 — the First Fruits of Its Multi-Billion-Dollar Meta Superintelligence Labs (MSL) Effort

(2) SpaceXAI Releases Grok 4.5 — with Technologies from the $60 Billion Cursor Acquisition

(3) Anthropic & OpenAI Roll Out AI Extras Alongside Fable 5 and GPT-5.6

MP OVERALL TAKE

Gadget AI — Samsung Readies Its Wide Foldable + Wearables for July 22

Questions

Q1 — What’s Michael’s primary appeal in foldable phones?

Q2 — What’s Michael’s biggest criticism of the foldables?

Wrapping up

Full Source Reading —

Event 1 — Meta’s Muse Spark 1.1 (MSL)

Event 2 — SpaceXAI’s Grok 4.5 (+ Cursor)

Event 3 — Anthropic & OpenAI’s Extras (Fable 5 / GPT-5.6)

Gadget AI — Samsung July 22 Wide Foldable + Wearables

Clips from today

Clip 1 — Apple’s Foldable Phone: A High-End Novelty

Clip 2 — AI’s ‘Parrot Can Talk’ Moment

Clip 3 — Meta’s Muse Spark 1.1 vs the Top AI Models

Clip 4 — AI Browsers: OpenAI Steps Back

Clip 5 — AI Models: Early Days, High Costs

About AI Ramblings Daily (ARD), and AI-RTZ

Links used in today’s show (already embedded inline above; listed here for reference)

Want the latest?

More like this

AI: Weekly Summary

AI: Apple's new 'All Clear' approach across device/OS ecosystem at WWDC 2025. RTZ #746

The Bessent Pivot. Inflate, Default, or Build?

Let’s be friends!