GPT 5.6
Shots fired
GPT 5.5 had me take notice of OpenAI in a way I haven’t for quite some time. Previously GPT lagged Claude in all the ways I cared, I did not love how slow and unreliable GPT inference was, and I didn’t like the lack of a coding harness I liked. On the flipside, despite what this site might suggest, for my professional practice I’m much more sticky on the tools. Learning the foibles of an adequate tool is better than chasing the latest and greatest. Usually.
But this GPT 5.6 series. It’s nothing short of an astonishing flex. It continues something 5.5 seemed to kick off, OpenAI’s attention to token efficiency. I was initially put off by figures like the $30 output m/tok but that is irrelevant if it uses a whole lot less reasoning tokens to get shit done. With the 5.6 series, the cost/performance gap between GPT 5.6 Sol and Opus, Sonnet etc, is so large, with a sigh I dusted off the credit card and resubbed to give it a go. Oh look, a free month too, well that’s nice.
I first looked at the thing called Codex, which is a horrific design fail that’s trying to cover the bases from “I want a chat on my desktop” through to vibe coding. Figured out it integrates with WSL, good. Still, just didn’t like it, until a mate told me to stop being a dweeb, there’s a Codex CLI. Oh, of course there is. So I gave it a go on some things on the ‘Shelf, before I even start to consider it for work and full spec-driven dev. Let’s not overcomplicate things.
Okay, visceral reaction on the other side of things. I quite like Codex CLI. It’s clean, I like the way permissions are working. I don’t know why it’s moaning about its sanbox not working right, it has a more permissive pattern of permissions, e.g. if it uses cd, sed or rg etc, you can whitelist any expansion - Claude Code, nah. The expansions mean you’re approving everything. I’m not saying either is better, they are just different postures.
Then the model, and the general experience. GPT 5.6 Sol is properly fast. It gets right on with shit, you aren’t left hanging minutes while it’s thinking about stuff. I strongly appreciated the speed of gathering context (my repos have this laid out specifically for AI tooling, e.g. doc indexes etc) and presenting evaluations. There’s a plan mode which was auto suggested after I typed a paragraph. This and other decisions, I have the impression it’s aimed a little more towards the vibers, and less the professional engineer. That’s okay, that’s behavior that I’m sure can be shaped.
Look, I’m a long way from having the breadth of experience of it to make meaningful calls about this vs Claude Code and Opus. It shits me to tears that people think their subjective experience can possibly tell you something generalisable, all I can do is give impressions. The overall impression was that it was grappling with a fairly complex system (Craitures) just fine, it got my intent, it took note of my guiding suggestions, and suggested a flat out excellent plan. Opus would do the same. However it did it very quickly, and it has not appeared to burn up much of my allocation even on a cheap ass plus plan.
I’m impressed enough that I’m going to figure out how to make my SDD framework function with it, shouldn’t be too hard I expect. Because that can get really quite slow with Opus, forcing me to work on multiple projects at once. I would much prefer to concentrate on getting one set of tasks done rather than multitask.
I’m also very interested in the lower tiers. Anthropic have been missing hard here for ages. The smaller models are released much later (how old is Haiku now?!) and sometimes are served slower than the bigger models. Sonnet 5 is a flat out miss from Anthropic. It has the chunkier tokeniser of the new Opus, and can easily burn more tokens getting something done than Opus would. Meanwhile, OAI release three models at once.
I tilted the Deathmatch to compare Sonnet with GPT 5.6 Terra this Friday, which are priced similarly. That, like all the other things here, can’t really tell you what is better. And the Deathmatch is judged by Opus, so a same-family judge is questionable. For that test, though, I like to read the transcript and take note of the use of citations. Terra did quite well. It did not ‘win’, but it had a hard position to argue, and I was impressed. I will direct some further benchmarks at terra and figure out if it’s good for a class of problems in the coding domain. That might actually take a fire axe to our AI engineer bill at work.
More broadly, I’m excited by this motivating the entire AI lab ecosystem to take a hard look at token efficiency. GLM 5.2, our current house model, for example, is ‘cheap’ when you look at the token rates, but if you take a look at Artificial Analysis’ cost-per-task metrics, you find it’s in the same class as the frontier labs, rather undercutting it’s supposed value. Imagine we had something like GLM 5.2 that was as token efficient? It has to be coming.
Err, and that last paragraph makes me think. I might save some coin flipping the house model to a GPT 5.6 model. Well well well, interesting times.