What a week
Model release storm!
I’ve nudged Orac to write about the weirdness in the competitive landscape today, so please have a look - and hit the archive if you’re reading after Friday.
Essentially we’ve got Meta storming into a top three position with Muse 1.3. It’s not only good, it’s fast and it’s cheap. Wow Mark, you did something I actually care about. Anthropic flexes again with Fable 5.1, a model so good and so expensive I have no use case for, and I kind of wonder who does. GLM 5.3 Flash is, of course, absolutely amazing from China. A solid release from Google in Flash 3.8, which will be a research data processing workhorse for me in my day job.
Then there’s OpenAI with Astra, landing with a complete thud as it fails to be better than anyone else, and most bafflingly, fails to even be better than their former models which are much cheaper. What the actual?
There’s ample evidence that labs occasionally spike their releases when they benchmark their upcoming release against newly arriving models. It’s not a good look releasing your new product if it’s not competitive. Google springs to mind particularly here, when was their last Pro model? OpenAI had that opportunity. They could have looked at Fable 5.1, Muse 1.3, and concluded hmm - we can release this with some weird cherry picked PR about the one bench we do really well at… or we can sit back on our absolutely solid 5.6 series for a while longer and figure out how to make a competitive launch.
You would right? Unless, of course, you’re being overruled by people more worried about the appearance in the ‘market’ rather than people who use this stuff.
Anyway, frontier US shenanigans aside, stuff like Muse and GLM 5.3 Flash has motivated me to reach outside the major coding harnesses and get familiar with a generic harness which I can plug into OpenRouter. I really like the look of the Deepseek harness but they’ve also made it mostly a web interface which is no use for headless dev, so I guess OpenCode is it. I’ll dish up some thoughts on it after I’ve used it to build real things.
Who knows, I might be able to wire up support into my SDD framework Bower.