AI Models thread

[mid_gen]

*Another suggestion for an AI subforum while I'm here. You can't search 'AI' because it's too short. So many discussions scattered across different subforums.....

Specific thread for the latest AI model releases.

Today Mistral Large 4 has been released. It's nice to have a European option, and open-weights too. Will be giving it a whirl later today.

My daily drivers at the moment are Claude Opus 5.5 and Fable 5.1 depending on the work. I'm also using Deepseek v4.1 Flash as a summariser for some automations.

I ensure my pipelines are all compatible with Claude Code and Opencode so I've always got a route to switch models and toolchains. I may configure a mistral agent that I can call from Claude Code to adversarial review things.

 

[mid_gen]

I feel like I should try out some other models and settings, but I just use Opus 5.5 on medium for everything. It seems to just work for the things I want it for without chewing through my subscription usage.

Just bear in mind Claude Max won't be subsidised the way it is now forever.

I rolled my own activity/issue tracking system on Monday, with CC and Opencode plugins, so it tracks all token usage across all providers.

I'm burning about $40-50 a day on Opus at API rates.

This is why I continually evaluate various other models and providers and make sure I can switch as and when I need to. Deepseek V4 flash was the first one that I really felt was capable and affordable enough to use as a daily driver coding agent at API rates. v4.1 and GLM 5.3 flash have pushed the boundaries even further....and the distils will keep getting better.

For now though Opus 5.5 and Claude Max are a luxury I will make the most of. Although I do keep meaning to try Codex again...I do have to focus my time and effort on my business, not evaluating LLMs....and Opus 5.5/Fable just gets **** done really fast and is very affordable still.

 

[mid_gen]

I've been hitting limits on a 5x max plan each week for the last few weeks.

All based off 1 Claude Opus 5.5 ultracode review to improve how my game can become beautiful (Gettin Opus 5.5 to carry out the work on Medium). It's scheduled about 150 hours of agentic workload which I'm chipping away at and am very interested to see what it can output. As a non-coder it blows my mind.

Got it doing its own dev blog here: https://pixelpunk-journal.pages.dev/

Ultracode is way overkill for a task like that, you're just demolishing your usage limits for little tangible benefit. Ask Fable to do a review (not ultracode) for a plan, implement with Opus 5.5.

 

[mid_gen]

Have you all been using AI for orchestration on Claude as well? I've been using Gemini for a few projects recently (more in terms of managing large document sets, not coding) but the memory and lack of context can be quite limiting in terms of overall use. On the plus side, for ~£9 per month I'm getting a lot of usage out of 3.8 Flash and only ran out of tokens once in a few months.

Most people using Claude will be using Claude Code as their harness, so Opus 5.5 is the default model for all orchestration and subagents, although you can change to whatever is suitable for your task.

 

[mid_gen]

Haiku 5.5 out today. About time because Haiku 4.5 has been obsolete for months. Can't see myself moving away from deepseek v4.1 flash for summarising work though.

 

[mid_gen]

I've been using Claude for a couple of months now and it works well for my needs. I don't entirely trust it write large amounts of code for me though. I prefer for it to tell me what I need to do and then implement the code myself.

I don't know how anyone is wasting time writing code by hand anymore, you might as well go back to making punch cards if you're going to be that archaic.

Are you not using Claude Code? The Claude chat app?

If not, then you really need Claude Code, if you think having Claude tell you what to do is impressive, wait to CC starts smashing out stuff for you. I'm a 20+ year professional software developer, and I trust CC to write excellent code these days. Even just a year ago it needed a lot of handholding, but in the last six months it's reached a tipping point where there's simply no reason to write code by hand anymore.

 

[mid_gen]

On a serious note I use a lot of tricks which AI has no ability to do as things stand (and/or I'd have to get the agent to go back in and implement in every specific case where it would be useful and if doing something to optimise code at a global level have to keep track of that manually anyhow) and/or do really esoteric things like misusing Raw Input to interface with an ATtiny85 as a HID device to work around the lack of support for a serial USB interface, etc. (which AI has no clue about).

Oh I'm using all the little performance tricks and stuff I've learned throughout my career. I've included it all in the agent instructions so it doesn't need specific prompting, it just implements features exactly as I would do it myself.

Literally just a few months ago I would be watching it like a hawk and frequently jumping in to change things and correct it. It's partly down to just learning how to use them effectively, partly down to the models improving, but they have been 'good enough' for a few months now.

I also enjoy the process of working up code and find quiet satisfaction in really nailing a neat optimised implementation.

The craft of software engineering is satisfying for sure, but I find much more satisfaction in just being able to implement things I would never have dreamed of allocating the time and resources to do previously. I built some systems yesterday that would have taken months if I was working by hand.

In a way I'm lucky that my trade is game development, because games are fundamentally an emotional product, the aim is to put the software in front of a player and have them feel some emotions, and LLMs will never be able to do that part. The crafting of a gameplay experience is a uniquely human skill. If I was just producing enterprise software in this environment I would be pretty uninspired with the career outlook tbh!

 

[mid_gen]

I don't know how you can maintain and improve your skills or keep up with the shape of your codebase if you're never writing code.

I didn't say I never *read* code. I'm frequently discussing the shape of the codebase and doing regular ongoing maintenance, refactors, improvements.

Using coding agents and vibe coding are not the same thing.

 

[mid_gen]

Games are going to be one of the hardest areas for AI to really replace a good human coder or be anything more than an assistance tool especially with breaking new ground with technical innovations, though as I've mentioned before I'd turn to AI every time when doing stuff like writing mundane shaders for the 1000th time.

I'm curious if AI could ever do what I did recently with taking the Quake 2 RTX source code and adding a mouse driven UI implementation in purely server code interfacing with and massively misusing the HUD drawing routines in the client, without touching any client code - even if I only did that out of curiosity if it could be possible rather than any practical application. There are so many hurdles to overcome in terms of understanding and seeing where you can abuse the message protocol used to send the HUD layout, working around the update rate of the network and HUD layout update rate, misusing the player's 3D view angles to translate it back into something useable for the 2D screen while also locking the 3D view updates, etc. etc. sure some aspects of that you could use AI to assist with but I highly doubt any model any time soon could understand all those challenges and how to overcome them.

I think you’d be surprised what frontier models can do now, even given the constraints. They’re well able to understand 3D space and come up with genuinely novel solutions. Fable 5.1 rooted out some quite obscure edge case bugs in my 3D navigation system I’d never noticed.

 

[mid_gen]

Unfortunately it is difficult to explain the challenges to anyone who hasn't spent extensive time with modding Quake 2 or working on a game using the underlying engine but it requires quite some out of the box thinking to work around the limitations and some pretty intimate knowledge of the engine itself.

That’s all irrelevant if you just give it the source code, all the information about the problem space it needs.

 

[mid_gen]

If you had a good enough understanding of the problem and engine limitations then you could iteratively use an AI agent to produce the component parts, but you couldn't just give it a high level prompt and get anything useful, this is a problem especially with game engines and especially older games.

For example if you fed it the full source code and said "implement a radar showing the position of hostile NPCs" it could probably go away and do that, if you said (and there are reasons to do this) "implement a radar showing the position of hostile NPCs using only the gamex86/base server code" it would likely fall flat on its face and produce completely unworkable code (or something really hideous if it partially cracked it).

EDIT: And I have no idea of the breadth of your experience with different game engines but I think a lot of game developers' minds would be blown if they had to spend awhile implementing modern game features within the scope of the original limitations of the Quake 2 engine and the problem solving it requires heh.

I’m not trying to downplay the challenges of working within limitations enforced by technology or engines at all, been and done plenty of it, and I agree it is probably the most interesting and rewarding aspect of software development

But I think you would be surprised (and depressed ) at what frontier models can do now. Even as opposed to just a few months ago.

I might set a model trying to solve your challenge later if I have some time

 

[mid_gen]

Are any LLM's as good as Claude Code? I've been using QWEN 3:14b and also QWEN 3.8:27b (bit of a struggle on a 5070 Ti) when feel 3 isn't good enough. Both struggled at a particular task, getting it wrong multiple times and not implementing a simple rule. I gave the code to Claude (Sonnet 5.5 I think) and it got it right first time. Same code, same prompt.
I'm still experimenting, and it might be in this due to the language being less widely used, but is there anything as good as or closed to the closed source models?

The big models like Claude Opus are upwards of 2 trillion parameters, so they are always going be in a different league to small open weights like Qwen 3.8 with only 27 billion parameters.

I have tinkered with the smaller models and they are fine for doing busywork, going and doing specific tasks that they've been clearly instructed on.....but for more open-ended tasks, troubleshooting, bug fixing, architectural work, the bigger models are more or less required.

You can get full-fat 2 trillion parameters open weight models, like GLM, Deepseek, but you need an absolute monster machine to run them (750GB+ VRAM).

There are some increasingly good open weight distils like GLM 5.3 flash, Deepseek V4.1 Flash that are very close to Claude Opus, OpenAI Sol in terms of their capabilities, but at a fraction of the price. I use these through Openrouter for doing adversarial reviews of Claude's work.

It's quite handy to use Openrouter to try out different models, I always have a play with the latest ones as they come out. At some point Anthropic are gonna start charging sensible prices for Claude and we'll have to adapt or switch. My custom tracking software tells me I've used $800 worth of Claude in the past two weeks on my £90/month plan

 

[mid_gen]

Bit of a loss leader pricing model at the moment then. Will be interesting when people's bills go from £90/month (1080/year) to £1200/month (14,400/year).

Well at the moment it's very much a case of making hay while the sun is shining, and just making the most of the subsidised inference. I don't need to be using Opus for everything, but I might as well as long as I have the usage.

This is why I experiment with lighter models, to make sure I can switch to full API rate pricing when I need to. Deepseek V4 Flash was definitely a turning point where I felt I could get by on API pricing when I need to, and they are only going to get better.

 

[mid_gen]

Are you using a harness? Quite a large part of what makes Claude Code good is the "Claude Code" bit rather than the model itself. The stuff which will run on my 3070 just isn't capable of the job so I haven't experimented much myself but from what I understand Open Code is decent. None of the local stuff really reaches the level of Claude Code (or rivals) but it has the advantage of cost and control, and put into the right harness it can do good work for you especially when you're AFK and just leave it cooking -- at least that's what I see from other's online.

Opencode is good, in fact I prefer the presentation to Claude Code.

Everyone using Claude Code should also setup Opencode and Openrouter....firstly so you have a backup if Claude is down, secondly so you can easily try out new models and be ready for teh usage-pocalypse

 

[mid_gen]

There are so many tools, configurations out there that's it's a bit overwhelming. Where do I start? What setup should I have, which tools?
I've been using Ollama & Aider and run aider from a batch file so I can switch between models (one batch file per model I use).

OpenCode + Openrouter is a good start if you don’t want to pay for Claude Code. There are usually some free models you can try out.

 

[mid_gen]

Done some rejiggling on my CI server and made some existing jobs run on the intel IGPU, and having a go at running Qwen 3.8 coder with Strata. I don't expect it to be terribly good on a 9060XT but I'm interested in seeing how these simple local models can start doing some donkey work (i.e. driving my playtest sessions).