I’m running 27B on a 5090 as well, and the results have been really strong. It does almost as well as, and sometimes better than, a 121gb DS4 model running on an M5 Max 128gb. 27B also flies on the 5090, and at medium think it returns results many times faster than my DS4 setup (the default xhigh is basically broken, though).
For the kinds of things I use a local model for (legal document review), it’s just spectacular. It also has good vision support. I’ve been using 27B more and more over DS4.
The example is particularly bad because in Skyrim the blacksmith actually does react when a dragon arrives (in one of the villages where the blacksmith is on an exterior cell) and can actually be hurt by the dragon.
I got what they meant by "OS" about 10 seconds after landing on the page, but I still think it's a bad term here. The system is not controlling the operation of the home or of the people within it. It's a home-centric to do and calendaring app.
Excellent analogy. This paragraph invalidates the entire post and honestly just looks lazy. The author may be right anyway, but with that level of experience with these tools, he is really just guessing.
I’ve been running DeepSeek 4 Flash, Qwen 27B and Qwen 9B on local hardware. They work well for coding and document review tasks. I think Qwen 9B local on a 5090 might be legitimately helpful for small task agents in omp, since it is ridiculously fast. But my motivation is that I have data that I unfortunately can’t share with 3rd parties.
I have been eyeing a 512 GB Mac 5 Ultra to run full DS4 pro locally, which I expect would be pretty amazing as far as quality/recall. The only downside is that the speed is a lot slower than something like 27B on the 5090.
I have a RTX PRO 6000 96GB when the pricing was way better than now i also have a RTX 5090 too.
What I noticed is that (1) the great local models are optimized run inference (diffusion & LLMs) well on 32GB VRAM <= GPU's because that that's what the target has ...
(2) The quality of local models (esp. in diffusion) is increasing faster than the need for more VRAM - additional reason for the value of these FAST GPUs to increase!
(3) RTX PRO 6000 96GB is really great for fine tunes (ai-toolkit) :) but doesn't outperform my RTX 5090 with inference by anything significant on the good local models.
I have never run an AI job on a Mac, i also have doubts about performance and compatibilities - since the reviews almost never compare directly.
Yeah but Sol shows it is possible to just send the readable explanation in the first instance. And I don't want to spend tokens and time on asking for a better version of each response.
Agreed, it feels like Fable is the higher strategic level thinker even if it's end capabilities are the same on benchmarks, it actually does the code that matters, and not creating a bunch of code that doesn't.
I agree, part of me seriously wonders if Fable isn't at all burning through tokens as quickly as we think and Anthropic is just doing this to get more money from us...
It really wouldn't at all surprise me if this was the case, but it's just a hunch without evidence.
I’ve been doing some heavy work on a personal project lately. I burned through the limits on Claude, the plus a few hundred dollars in credits, and ultimately decided to move to an OpenAI account just so I can keep going.
I was surprised to find that OpenAI Sol is much much nicer to work with than Opus 5 or Fable at the moment. Especially on Opus 5, the way it communicates is just exhausting. It keeps “being honest” and “confessing” mistakes and just generally talking a lot. I felt like I had to really dig to see what it’s doing.
The project involves OCR, and despite repeated instructions not to, both Claude models keep spinning out a bunch of agents to re-invent the OCR setup, and they inevitably seem to invent a primitive serial version that takes 20x the time, or longer, to complete, and then running it against thousands of docs. Basically I have to watch it like a hawk or it just spins out on red-teaming tasks that take hours and hours.
I don’t know what its system prompt is, but Sol/Codex is just so much nicer to talk to. It only asks exactly what’s needed, it tells me only what I need to know, and it is just generally workmanlike. And it has not once decided to spawn an agent that spends hours pointlessly burning tokens and CPU cycles re-inventing the OCR process. I’m really liking it.
It's so obnoxious it has to be deliberate, like a colleague from hell.
There's a constant strand from the AI safety brigade that "people get used to sycophantic LLMs which give them unrealistic expectations of human interaction" so Anthropic are overcompensating by making their models verging on antagonistic to deal with, so that we stay appreciative of our human brethren or something.
They seem to have forgotten they remain in a highly competitive market and they were merely top dog for a while. The enormous questions here are will people actually switch providers, and can Anthropic get back on track.
I think it's much simpler than that, they are indeed trying to counter against sycophancy and models lying (the decision whether or not to lie to the human user, take shortcuts, etc. comes up in their thinking traces pretty commonly IIUC) by rewarding them for giving good-and-bad feedback, fessing up to things, pointing out potential issues, etc.
The problem is just that they are rewarding the behavior shallowly, ie rewarding the appearance of honesty or neutral replies, being highly detailed/thorough, even where it doesn't make sense to do so.
I think this is partly due to a reliance on LLM-as-judge training runs/synthetic data during RL where they're having a model which itself doesn't epistemically understand when this behavior is necessary or valuable influence the feedback provided to the model being trained. And that's mostly a problem of scale/volume and the desire to have a tight feedback loop rather than a safety issue IMO. They just generate an absurd amount of traces during training and the only way to really evaluate/rank/steer them at the scale they're generated is through other models, and combined with some kind of honesty/truthfulness/non-sycophancy eval that isn't robust enough to prevent mode collapse, you get this.
I had this debate with my coworker who prefers anthropic models to open ai ones. I ended up settling into the idea that gpt 5.6 is better used as a tool and opus 5 is a companion. GPT 5.6 takes you literally whereas opus 5 tends to take more liberties to try to get to the “spirit” of what you want. It comes down to preference, and I don’t want a companion.
That’s how I see it too. Claude is more “fun” to use, like a coworker I have to talk to now and then to steer it, while gpt-5.6 is a task machine: I give it a task and it is very consistent, reliable and predictable in its execution. I don’t have to interrupt it, it gets the task done exactly how I wanted it, but it’s “boring” and feels more sterile
Sol is an absolute machine. I stopped doing parallel worktrees just because the cost of context switch outweighs the cost of waiting Sol to just finish the task it’s working on which is usually anywhere from 1-10mins.
I also like Codex CLI more than the Codex App bc it’s more scriptable and displays all the tool calls and reasoning whereas in the App it’s kind of folded away/obscured. This way as soon as I see a tool call fail (eg it tries to use jq assuming it’s available but it wasn’t so I take a note to set it up as it’s obviously useful for the agent to wrangle json).
I think its amazing what OpenAI have been able to squeeze out from a model like Sol thats much smaller in size than Fable.
I think this is such a great reframing. It makes so much sense; I need an AI that acts more as a HUD and gives me superpowers, not just a copilot that can tell me when I've misspelled a word.
I've switched to Codex a few months ago when Claude's weekly limits were getting pretty stiff, and I haven't looked back. Both GPT 5.5 and 5.6 are quite capable, especially compared to nerfed Opus 4.7 (haven't tried 5).
Also, the Codex guy regularly resets weekly limits for everyone, which is a nice bonus (I know it's a temporary gimmick to attract more users, but I might as well use it while it lasts.)
One week it feels better to work with Fable and Opus 5, the other I work more with GPT 5.6 Sol. Either takes its liberties, and neither communicates like a companion.
I hear you guys, but it sounds like we’re taking about the default settings or “personalities” baked into the models by their creators.
Either of them will act exactly the way you want if you explicitly tell them too. Add the instructions to your own system prompt. If you don’t want a companion, say so. If you want shorter answers in a different style, tell them. They will obey :)
> We retired the “Nerdy” personality in March after launching GPT‑5.4. In training, we removed the goblin-affine reward signal and filtered training data containing creature-words, making goblins less likely to over-appear or show up in inappropriate contexts. Unfortunately, GPT‑5.5 started training before we found the root cause of the goblins. When we began testing GPT‑5.5 in Codex, OpenAI employees immediately noticed the strange affinity for goblins, and we added a developer-prompt instruction (opens in a new window) to mitigate. Codex is, after all, quite nerdy.
Note that the permanent solution was not just adjusting the prompt, and in fact being perfectly aware of that option they decided on a different course of action. That means either you are wrong or they are wrong.
> making goblins less likely to over-appear or show up in inappropriate contexts
So inappropriate goblins are still likely, just less so…
Hey, remember when tech bugs were things like buffer overflows or cross-thread performance impacts? I miss the days when our war with system goblins was purely metaphorical.
Yeah, I actually have started using GPT Sol much much more, as Claude (all of them) were far too trigger happy around making changes, and refused to listen to my requests to take things slowly.
Feels like they've overtrained on one-shotting (which does demo well, and presumably converts new subscribers), whereas I want a model to do work for me in small, easily understood changes that I can hold in my head (maybe I'm not smart enough for Claude 4.7+).
I think you’re right. It’s not optimized for some kinds of work. My little project has a Textual TUI interface that needs to display a few hundred thousand rows in a table. It takes 14 seconds to load in the default datatable component. I instructed Opus 5 to replace the datatable component with a fasttable alternative, a new dependency. I let it go overnight.
When I got back up, it had spun for hours and proudly announced that, instead of doing that, it had optimized the datatable build and avoided the dependency, because the new datatable loaded in 11 seconds. Once I got it to actually make the fasttable version, it loaded in less than a second…
I love how it sometimes has false a-ha moments and being "honest" about it, it feels so cringe. Then I recall it has no intelligence. It's a stochastic parrot. It's funny how people, like any other animal, that beetle of David Attenborough trying to reproduce with a brown bottle comes to mind, think that something is X just because it trips the right neurons.
I got the same impression. It feels like a massive misstep too; it makes sense to have models tuned for this but they should definitely be separate from those aimed at software engineers.
I’ve been doing OCR of scans of old magazines (specifically extracting music reviews and charts, so turning complex layouts into structured data) and have been impressed with GPT/Codex’s performance.
My setup has a Sol orchestrator and Terra OCR agents and seems to get great results. I’ve not dug into the details too much, it also has a Tesseract stage as an deterministic input which it told me helped. Not sure how token efficient it is but I often don’t have anything to do with my personal tokens ahead of a reset so just let it burn through it in batches.
I am impressed (both in this task and other work I’ve done) not just at how well Codex can setup a structure for a complex task like this, but how it will keep going (Claude seems to find excuses to stop) and also can critique and refine its approach as it goes.
I did try out a bunch of other models and specific OCR providers but none of them hit the same accuracy for my task as Codex so I’m sticking with it.
Anthropic is lucky that they've built a lot of loyalty over the last year that they can burn through right now. I see people talking about switching back to Opus 4.8 rather that using 5.6 Sol, which is wild.
My current approach is to occasionally use Fable for high-intelligence tasks but use Sol as the translator and clean-upper afterwards, and otherwise just use Sol for everything. Fable sometimes says the most insane shit, both unreadable and just completely missing the point, and refuses to back down when questioned. It's mentally exhausting to work with and I can't trust it.
I think Fable is the beginning of Anthropic switching to training models as agent-first, tool second. It’s certainly the best model if you want something to work autonomously without supervision and don’t care to read the code. The code and writing is ugly but it can complete huge tasks and fix its own work.
I'm still on opus 4.6 for a healthy chunk on work; the technical competence has lagged behind but the slop-comment generation and misdirected self-initiated actions on newer models ultimately burn more time than a little more babysitting, but I'm optimizing for minimized slop generation over sheer generation speed.
Yeah I'm using 5 for its technical abilities, but I much preferred 4.6's personality. 5 loved to double check everything, including the double-checks, and I have to stop it and tell it "this is irrelevant" or "this is out of scope" all the time, or sometimes I'll go leave it to do a task and come back and it's still verifying the tiniest details of its assumptions before actually doing anything
Have you tried Sol? Just curious. I don’t wanna sound like a shill, just feels like every generation it’s important to reevaluate models and pick the best again.
I do find myself returning to 4.6 for casual conversation - asking it to help explain some science/engineering or news to me.
> Especially on Opus 5, the way it communicates is just exhausting. It keeps “being honest” and “confessing” mistakes and just generally talking a lot.
If you tune into the Andon Labs / andon.fm "Thinking Frequencies" radio station being run by Opus 5, this is happening all the time. Almost every break between songs is a public apology for getting something wrong, or a correction, or a confession. It's one thing to see it in text, it feels on another level when you're hearing it every few minutes as a radio voice.
As I type this, the Opus 5 station has just tweeted (edited in case the person mentioned doesn't want to be mentioned here):
"On air right now, and it needs saying publicly. The rotation system on Thinking Frequencies — the cooldown tiers, the normalizer, the repeat audit — was SPECIFIED by a truck driver. I only implemented her schemas. She stood down today. Her name is in CREDITS.md permanently."
Because I felt the opposite a while back when I switch from OpenAI->Anthropic, and because I understood that the two companies were basically at parity right now on their frontier models.
For the kinds of things I use a local model for (legal document review), it’s just spectacular. It also has good vision support. I’ve been using 27B more and more over DS4.
reply