I’ve been using LLM’s for work for so long that it’s easy to forget many people aren’t inside the bubble. Even people in software development aren’t all riding the tiger as it were, so I thought I’d share thoughts on my daily drivers.
OpenAI’s Lineup
OpenAI is pushing the best models since Anthropic took a bad turn after Opus 4.5. Therefore, it gets top billing here.
GPT-5.6-Sol
This is the current flagship model from OpenAI. This is your galaxy-brained friend who has a supercomputer built into their table for reasons unknown. This model speaks a terse form of English that indicates it hasn’t seen the sun for a while. Hold off on the invite to the dinner party. However, if you’re looking to implement something big and important in the codebase, Sol’s your model. Prone to over-engineering so shoot some thinly-veiled insults at it to rein it in before you let it rip.
GPT-5.6-Terra
The mid-level model. This is a great workhorse, this is the Senior dev that just needs a little more seasoning before ascension to full Wizard status. Work with Sol on architecture, and you can confidently tell Terra to carry it out. Cruise the diff and if you spot anything you don’t like, tell it what to do and it’ll see the error of its ways and fix accordingly. Strong on its own for most tasks, takes less time, but doesn’t move civilization up the Kardashev scale on its own.
GPT-5.6-Luna
The trusty and eager junior. Luna is lightning fast which makes it fun. Will get you the context you need to feed to its bigger cousins, or if you just want something explained. Very capable for any simple task.
Anthropic’s Lineup
I’ve majorly soured on Anthropic’s output as of late with the exception of Fable. Opus and Sonnet 5 are just disappointing. One saving grace for Opus is it excels at reviewing code. Maybe the only leg up Anthropic has at the moment.
Fable 5
The flagship model and Anthropic’s saving grace at the moment. I’ll save my distaste for this company for another time. This model is fantastic at writing and as strong as GPT-5.6-Sol on the technical front. It feels the most ‘alive’, has the most personality, although that personality is tinged with arrogance (not unlike some Wizards you encounter). Don’t ask it something you should ask smaller models unless you want a response dripping with contempt. By the way, you probably can’t afford it.
Opus 5
Yuck. I don’t know what happened here but it’s just kinda bad. Extremely verbose without justification, with a tendency to miss important details. Its verbosity feels like insecurity. Don’t get me wrong, you can get good work done with it, but it’ll feel like a trip to the dentist.
Sonnet 5
Serviceable but unimpressive. Don’t see any advantage over Luna.
The Rest
Grok 4.5/4.6
I’ve used these a bit out of curiosity based on the recent benchmark hype. The output is always quite good, but not enough to make me switch. My tinfoil-hat theory is that when Elon booked a compute deal with Anthropic, he distilled their models into Grok 4.5. Grok also bears attention because Elon has endless compute, and the Cursor team (who Elon bought) is very legit.
Google Gemini
I lost faith in Gemini in the v2 era when I took it for a thorough drive and it failed miserably. I haven’t tried their flagship models since, but when I have a massive blob, I use the 3.7 Flash model to pick needles out of the haystack and it is good and fast at that. I think it’s totally serviceable for utility-tasks, and writing simple messages.
Haven’t tried but making waves
Kimi and GLM get good reviews and score really well on benchmarks. I have never directly used either of these models so I can’t speak to them but they bear mentioning. These models represent a scary proposition for OAI/Anthropic (and their investors), because they seem to be quite close in quality at drastically cheaper prices.
Conclusion
OpenAI is still king, for the time being.
If you’re made of gold bars, use Fable. Opus and Sonnet are frustrating.
Grok/Kimi/GLM are legitimate contenders.
Google needs to figure some things out.
Important caveat, the harness an AI model gets used in is extremely important. I’ll go into detail in a future post. You wouldn’t believe the effect it has on model performance, which is something to be aware of!
Leave a Reply