Ryan’s Model Guide: September 2026

Ryan published his September 2026 model guide, ranking coding models across Codex, Claude, Z.ai, and Meta for tasks ranging from hard coding to computer automation. Ryan named Fable 5.1 as his pick for hard tasks and deep reviews, GLM 5.3 at high effort and Opus 5 for medium tasks, and Meta's Muse 1.3 plus Luna at xhigh for easy mode, while calling Codex's Terra "a hard no" for being "dumb" and relatively expensive. Ryan said he downgraded to Opus 5 for most serious work after exhausting tokens on a new project and pulled subscriptions from Z.ai and Meta, noting Muse 1.3 beat GLM 5.3 Flash with more usage than he could spend.

Ryan’s Model Guide: September 2026 I ran out of tokens pretty quickly this month new fun thing I’m building was a bit hungrier than expected , and so I had to pull out a couple of subs for providers I don’t normally use Z.ai and Meta and I was pleasantly surprised at how much work I could get done. I even had to downgrade to using Opus for most of my serious work and after some very painful cycles managed to make it not terrible. So here is Ryan’s guide for models September 2026 Codex - Astra: it is greedy and even on a Max 20 will chew tokens. Use it for computer automation where it shines, otherwise keep it on the side. Seriously one can configure Cloudflare, AWS and Railway in a browser and automate the custom TUI you built all at the same time, very impressive. - Sol: effective enough, I used to use this a lot on Xhigh, less effective for the hard stuff than Fable, so I would just rather use a cheaper model for most things - Terra: Personally this is a hard no as it is dumb, requires lots of guidance and is still relatively expensive. I stupidly use it for infra tasks and I always regret it. - Luna: scary how much OpenAI gives you for this. Set it on xhigh and guide it… it can do most coding tasks as long as you tell it to research first and to plan out what it is doing. It is a solid review bot. For the token spend it is really amazing. Claude: - Fable 5.1: Give it tasks you aren’t totally sure about, or just want to experiment with, use it for killer reviews but it is super greedy. I blew 10h of quota one time on a review and was waiting on another 5h limit before giving up and switching to a cheaper model. Never use it for easy stuff as it will make an overly complicated mess. - Opus 5: once you get past the incomprehensible language, and desire to just guess about facts confidently this is an extremely solid model. Tell it to read specific docs or code, don’t ever trust it to just do a thing without that instruction and make sure the verbiage has been toned down with some Claude.md entry telling it to avoid all the stuff you hate about it. - Sonnet 5: there are way better models for cheaper everywhere else. Z.ai Note I prefer using this with a codex custom profile will blog about this . Their Zcode env is pretty meh no worktrees or sandbox - GLM 5.3 - wonderful model especially on max but chews up tokens in their equivalent of 20 max very quickly. I use it with high effort for most anything that Opus can do and it is very effective. Had some big wins from this in the past month. - GLM 5.3 Flash - I use this a lot, you need to watch it closely but it runs quickly and produces solid code for simple tasks. Do not throw hard stuff at it; it will be very silly, but you get a lot of this in their sub, so use it. I use this at home with Hermes and I’m very happy. Meta: Muse 1.3 is better than GLM-5.3 flash and you get even more usage out of their sub, I can’t use all my credits Muse code isn’t great but it is improving rapidly. Ok so where does that leave me for my day to day: - Hard tasks - Fable 5.1 - Medium tasks - GLM -5.3/high, Opus 5 - Easy mode - Muse 1.3, Luna/xhigh - Deep review - Fable - Most Reviews - Luna/xhigh - Computer automation - Astra 6 This is all very unscientific but I figure this will be useful for many people who are just not sure what to use and more importantly for everyone as befuddled by Opus 5 as I was.