- Product Upfront AI
- Posts
- 🤫Six mysterious AI models just broke the internet
🤫Six mysterious AI models just broke the internet

Wanna know what it feels like to watch a frontier AI model get built in real-time?
Six mysterious models just appeared on LM Arena with code names like "Zenith" and "Summit", and the entire AI community is convinced we're getting a sneak peek at GPT-5 before it officially launches.
It's like OpenAI decided to build their next model in public, and honestly?
It's fascinating to watch. We dove deep into what each model does and why this approach might change AI development forever.
Here's what I’m going to share about AI today:
Six anonymous models are dominating coding benchmarks (and we think they're GPT-5)
These aren't separate models, they're probably "experts" that combine into one super-AI
Launch could happen as early as this week (ThursdAI, anyone?)
This hybrid reasoning approach might continue through GPT-8
The GPT-5 Models That Have Everyone Talking
Okay, y'all, something wild is happening on LM Arena right now.
Six anonymous AI models showed up out of nowhere and started crushing every benchmark.
We're talking about models with names like:
Zenith (the flagship - probably full GPT-5)
Summit (general-purpose reasoning powerhouse)
Lobster (likely GPT-5 mini for speed)
o3-alpha (pure coding specialist)
Starfish (possibly the open-source version)
Nectarine (the odd one out - might not even be OpenAI's)
so far, 5-6 mysterious models on LMArena are rumored to be from openAI:
• zenith
• summit
• lobster
• nectarine
• starfish
• o3-alphazenith, lobster, and o3-Alpha claim to be top-tier coding models
surely at least one of them has to be the full GPT-5
— Haider. (@slow_developer)
4:30 PM • Jul 26, 2025
And here's the kicker: AI researcher Ethan Mollick asked "Summit" to create a starship control panel in p5js (link)
The thing spit out 2,300+ lines of working code on the first try. Complete with warp drive, shields, and voice commands.
The performance claims are absolutely bonkers.
We're seeing reports that these models are outperforming Claude Sonnet 4 at coding tasks.
If you know Claude's reputation among developers, that's... saying something.
Why This "Mixture of Experts" Approach Changes Everything
These probably aren't six separate models competing against each other.
They're likely individual "experts" that will be combined into the final GPT-5.
Think of it like building a superhero team:
One expert crushes Python coding
Another writes beautiful, creative content
A third handles complex logical reasoning
Together? They become the ultimate AI
By testing them publicly on LM Arena, OpenAI gets massive real-world data to see which experts perform best before the final assembly.
It's brilliant, sneaky, and we are literally watching the sausage get made.
According to The Information, GPT-5 will be a "hybrid reasoning model" where you control how long it thinks before answering.
Like a dimmer switch for AI intelligence - crank it up for complex problems, dial it down for quick tasks.
When Are We Getting This Thing?
The launch timeline is all over the place:
Originally heard: End of July
Then: Early August
Now: July 31st rumours
Official leak to The Verge: August
But here's what we know for sure - these models just disappeared from LM Arena. Testing is done. Launch is imminent.
Most likely scenario? Thursday launch, because OpenAI loves their ThursdAI releases. But honestly, your guess is as good as ours at this point.
What This Means for You
If you're a developer: GPT-5's coding capabilities could revolutionise how you approach complex refactoring and debugging. We're talking about handling real-world programming tasks that currently make engineers pull their hair out.
If you're building AI products: The mixture-of-experts approach could reduce costs while improving performance. Specialised models might offer better results for specific use cases.
If you're in business: Advanced reasoning capabilities could automate more complex workflows. The hybrid reasoning model gives you more control over AI responses.
For everyone else: We're witnessing the future of AI development happen in real-time. The gap between AI capabilities and human tasks continues to narrow.
The Reality Check
Not everyone is buying the hype train.
AI critic Gary Marcus urges caution against runaway excitement.
His piece "What to expect when you're expecting GPT-5" is worth reading for some cold water on the whole situation.
Fair points: Public testing can feel underwhelming when the official launch happens.
Real performance might not match the leaked benchmarks. Integration challenges are real.
We're getting a front-row seat to frontier AI development for the first time.
And based on what we've seen so far? This is going to be remarkable.
And honestly? We can't wait to see what happens when all these mysterious pieces finally get put together.
Before You Go
Help me help more people:
I just shared my newsletter with a directory
It will take 30 seconds to leave a review.
But it helps me reach thousands more people who need to learn this stuff.
Your support means everything. Seriously
New to AI but curious about what's possible?
Subscribe here for weekly tutorials that actually make sense.
No jargon, no hype, just step-by-step guides you can follow.
Reply