HUGE news this week in the world of A.I. The top 2 are now a 4 way split. Google’s “Argon” was in preview and just dropped today, available to customers on certain plans. Meta’s new “Muse” Wonder Full model is right up there as well. Both of them are rated about equal 2nd place. It’s the 14th Oct, 2026.
What about coding, specifically? Glad you asked.
Here’s what that looks like, as the “across the map” image, below. Highlighted are Claude, Codex, Muse and Argon. Look at ’em all, all up there, all up the top.
Why this is seismic
See that graph? For all of time, it’s been black and brown trading blows, with a prior brief appearance from Google. This is the first time we’ve had a proper 4 way split, properly clustered. Water’s getting warm.
And what about cost? Are these game-changing, industry-disrupting bargains? No. Not really. And Claude’s on notice.
So what’s the sentiment?
I’m hearing: Claude for planning, Opus 5.5 for multi-agent workflows and fantastic one-shot decks and explainers, Sol 6.1 for farming out coding grunt work with incredible amounts of output for your usage tiers. Opinions are still coming in on the two new players.
Muse increased token efficiency over its earlier models, but not as much as the quantum leaps from OpenAI’s recent Sol. Still more efficient than Claude’s offerings.
Both Muse and Argon offer improved hallucination rates, which is, in my opinion, a quietly dark horse. That means less token use and fewer revisions. Don’t underestimate it. Muse has the second lowest hallucination rate, and Argon’s the new king. We’re talking 23% and 15%, compared to as high as 46% for Opus 5.5 and 43% for OpenAI’s Astra.
It should also be noted that the desktop apps and wider ecosystem functionality for the big boys are still supreme. ChatGPT’s image generator and gpt-live two-way realtime voice models are absolutely insane achievements, and remain kings.
It’s not all roses
Word is these are still slightly a generation behind in terms of “harness” and workflow. A bit wordy. A bit verbose. Perhaps using more tokens to do a job, which offsets the cheaper rates and drives the cost of those tasks up.
If that’s the case, they’re only a few steps behind the leading two, who both took big jumps in efficiency and verbosity in their latest models. I expect the others to catch up soon.
This isn’t a super deep dive. Check out the news feed and data over at Artificial Analysis for some incredibly detailed analysis and breakdowns.
There’s a lot more going on, though
All sorts of specialised models.
- Microsoft Decision 1: Microsoft released its JEV equivalent.
- Odyssey 3 is absolutely wild. Real-time rendering of adventures. Imagine a prompt that’s “You’re a golden retriever on a beach” and you can steer doggo around with WASD. Live. And multiplayer. Seriously insane, you have to go watch some demos on the site.
- Mistral Large 4: Mistral becomes the smartest AI outside of China and the US, again.
- Reflection Beam: open-weight models still on the march.
- LightOnOCR 3: lots of specialty models getting really good at what they do, like document extraction.
And far, far more.
Shout out to MLAI for hosting me today for co-working and mentor sessions at Stone & Chalk in Melbourne! More mentor sessions firing up shortly.
Originally published on LinkedIn.
Leave a Reply