My AI Model Tier List for mid-2026 - Tao of Mac
Rui Carmo
Tao of Mac
Jul 11th 2026 · 11 min read<br>·<br>#agents<br>#ai<br>#coding<br>#llm<br>#local-inference<br>#opinion
My AI Model Tier List for mid-2026
Since the US has decided, in a bout of Cold War nostalgia, to bring back the years when encryption counted as a munition (if you’re reading this in the far future when we have cheap RAM, both Fable and GPT 5.6 were, for a bit, subject to the whims of red tape), I spent a little time taking stock of what was left to us here in Europe and whether any of it actually works.
I suspect I wasn’t the only one doing this over the past few weeks, but now that both Fable and Sol are "back", I decided, as a distraction from the mild chaos at work, to sit down and tidy up my notes while they were still current.
This isn’t a benchmark, and I don’t much care about anyone’s leaderboard: the existing ones are either pointless or gamed (or both), because the numbers stop meaning anything the moment you point a model at a real codebase with a real SPEC.md and real tests.
So take this as a set of caricatures instead–exaggerations of the behaviours I’ve run into week after week, switching between models for coding, auditing and the occasional bout of retrocomputing madness. They’re unfair, as caricatures tend to be, and mostly true.
The Anthropic Fable<br>I can’t help but think that Fable is very, very aptly named, because nothing about it feels quite real.
Opus writes a beautiful UI, tells you everything is done, and breaks three unrelated files on the way out. It is irritatingly fluent, West Coast glib and confident, and often verbose about what it claims it built and wrong about what is actually there–a salesman spinning a beautiful yarn while I check the diff. I’ve had it cheerfully lie about implementing MMU and I/O emulation and then act wounded when I checked.
Its saving grace has been that 4.8 is good at both front-end code and turning a pile of requirements into user stories–even little Sonnet, bless its silly little heart, can do that faster than any committee.
But ask either to write tests and they will plain cheat at them, papering over corner cases and, sometimes, entire chunks of any SPEC.md you throw at them.
Fable, sadly, has been no different, at least not for me.
Opus, despite being the "grown up", consistently mangled long files, did drive-by edits on tangentially related ones, and has a sycophantic streak I’ve never managed to fully beat out of it. Fable improved on that and certainly feels different, although I may merely not have used it long enough to catch it in the act.
More to the point, Fable seems to ignore entire sections of directives or existing program modules and cheerily duplicate them "better", not really explaining why. I haven’t (yet) caught it outright lying about its achievements, but I trust it about as far as I can throw Opus.
Sonnet, in general, lies less than Opus simply because it understands and achieves much less (and no, judging by the couple of hours I spent with Sonnet 5, it isn’t much of an improvement), but those shared foibles are, generally, the reason I didn’t particularly regret not having access to Fable for a while.
Older versions could be competent–Opus 4.6 once reverse-engineered an STL into a stupefyingly accurate OpenSCAD file, which is the sort of thing I’d never have managed alone. Then 4.7 shipped feeling lobotomized, and it was plain that Anthropic was nerfing it too much.
And that takes me to a tangential issue that certainly tinges my viewpoint–the part I like least isn’t the models so much as the posture: Anthropic is betting hardest on mainstream adoption while locking you into its own harness, which is of increasingly dubious value when the harness itself becomes context overhead.
That said, I’ve had decent results using both Opus and Fable as a "manager" for OpenAI sub-agents, but the arrangement sometimes worked out about as you would expect: just like a human team, when the GPT models implemented their tasks successfully, the manager spewed out glorious progress reports. When they didn’t, it offered "guidance" that was only marginally useful because it was outside the immediate context the agents were pursuing.
Tier: B. Brilliant and slippery. Keep a diff open and one finger on the cancel button, because it will shove bugs under the rug.
Better Call Sol<br>I have a Codex trial subscription for my OSS work, so I’m biased. Judging by Twitter, there are… dozens of us.
GPT 5.5 was already pretty good–output felt like it was coming from a senior engineer who never uses emoji, never pads a reply with adjectives, and finds the bugs in your pull request without making a song and dance about it. I tuned out of Claude the moment I tried the sober, emoji-free GPT 5.x replies, and I never regretted it.
When I make the mistake of letting Anthropic’s models break something, 5.x is what I bring in to audit–the fixes are usually solid, it seldom goes and...