Which LLM and software plan is the most accurate in Metin2 development process?

[JackSbirrow 0]

Hello. I'm using some files and bootstrapped my local server. I'm a software engineer in real life, but developing on metin2 has always been so hard to me. That's why I thought having a try with AI in 2026 would be incredible.

I'm using Cursor with Composer 2.5, but it struggles a lot during development, lots of iteration.
Then I spent 15€ on OpenRouter to try Claude Opus 4.8 and it was insane. I asked to migrate proto system to database as single source of truth, do whatever it could and he did it first try. The problem is that Claude Code is very expensive, and I should need at least the 90€ plan.

Do you have any suggestions or comparison between models?

[hvi 45]

https://llm-stats.com/benchmarks/swe-bench-pro
This benchmark is aimed at measuring coding through agentic workflows (ie: how AIs are able to leverage the tools that different harnesses provide them)

Personally I'm using Kimi 2.7 and Deepseek V4 Flash, depending on the scope.
I tried Kimi 2.6 and GLM 5.1 aswell and was getting good results. GLM 5.2 just come out and seems promising, didn't have a chance to try it yet.
They're usually on par with Claude, in between Sonnet and Opus.

Edited June 18 by hvi

[JackSbirrow 0]

8 minutes ago, hvi said:

https://llm-stats.com/benchmarks/swe-bench-pro
This benchmark is aimed at measuring coding through agentic workflows (ie: how AIs are able to leverage the tools that different harnesses provide them)

Personally I'm using Kimi 2.7 and Deepseek V4 Flash, depending on the scope.
I tried Kimi 2.6 and GLM 5.1 aswell and was getting good results. GLM 5.2 just come out and seems promising, didn't have a chance to try it yet.
They're usually on par with Claude, in between Sonnet and Opus.

Are these models capable of handling architecture refactoring on these files? 
I used deepseek within cursor but it was so dumb honestly, I don't know if it got stucked due to Cursor limits, maybe I'll try on OpenCode. What about the other models? Are they free to use locally or need something like OpenRouter?

[hvi 45]

59 minutes ago, JackSbirrow said:

Are these models capable of handling architecture refactoring on these files? 
I used deepseek within cursor but it was so dumb honestly, I don't know if it got stucked due to Cursor limits, maybe I'll try on OpenCode. What about the other models? Are they free to use locally or need something like OpenRouter?

I don't let them refactor as I simply do not trust an AI with properly structuring the code. Not once I got a proper structure, not even with claude.
I use deepseek flash for bugfixes only, it's a relatively small model, I wouldn't trust it with bigger modifications. Didn't try the pro version which is substantially bigger.

Kimi and GLM are the real deal, both offer coding plans like Claude but with much higher usage limits at the same price points. Kimi is usually cheaper as GLM likes to spew out insane amount of tokens for reasoning.
https://www.kimi.com/code/
https://z.ai/subscribe

EDIT: For big refactoring the context window is probably your biggest bottleneck. GLM 5.2 offers 1M tokens context, so that's probably your best bet for your use case.
Minimax M3 should be able to handle it aswell, but I didn't have to chance to try it yet.

Edited June 18 by hvi

[JackSbirrow 0]

12 minutes ago, hvi said:

I don't let them refactor as I simply do not trust an AI with properly structuring the code. Not once I got a proper structure, not even with claude.
I use deepseek flash for bugfixes only, it's a relatively small model, I wouldn't trust it with bigger modifications. Didn't try the pro version which is substantially bigger.

Kimi and GLM are the real deal, both offer coding plans like Claude but with much higher usage limits at the same price points. Kimi is usually cheaper as GLM likes to spew out insane amount of tokens for reasoning.
https://www.kimi.com/code/
https://z.ai/subscribe

EDIT: For big refactoring the context window is probably your biggest bottleneck. GLM 5.2 offers 1M tokens context, so that's probably your best bet for your use case.
Minimax M3 should be able to handle it aswell, but I didn't have to chance to try it yet.

Yeah, my use case would be to simplify or modernize some core stuff like proto and database, before moving to bugfixing or implementing new stuff. I need something cheap as Composer but which can handle c++ large codebase like Opus. I'll have a look at those models.

[Matteo 69]

Claude Pro subscription (20$)
VSCode

I've created all the skills/agents/instructions with Opus (and when Fable was available, that updated and refined them) for the larger components of the game.
Then i use Sonnet for lightweight things, Opus for building P2P systems easily. You just have to prompt it well. 

If these are done, it won't consume as much tokens since it doesn't need to re-read and understand your whole infrastructure on every session.

 

[JackSbirrow 0]

30 minutes ago, Matteo said:

Claude Pro subscription (20$)
VSCode

I've created all the skills/agents/instructions with Opus (and when Fable was available, that updated and refined them) for the larger components of the game.
Then i use Sonnet for lightweight things, Opus for building P2P systems easily. You just have to prompt it well. 

If these are done, it won't consume as much tokens since it doesn't need to re-read and understand your whole infrastructure on every session.

 

Same, I used Fable to create skills. How are you handling it with the 20$ plan? Don't you get limits soon?

[Matteo 69]

1 minute ago, JackSbirrow said:

Same, I used Fable to create skills. How are you handling it with the 20$ plan? Don't you get limits soon?

No, i usually hit the 5 hour limit in like 1-2 hour of prompting. Sometimes its working for 40-50 minutes on its own, and still not limited.

[hvi 45]

2 hours ago, Matteo said:

No, i usually hit the 5 hour limit in like 1-2 hour of prompting. Sometimes its working for 40-50 minutes on its own, and still not limited.

That sounds like hitting limits to me

2 hours ago, Matteo said:

Claude Pro subscription (20$)
VSCode

I've created all the skills/agents/instructions with Opus (and when Fable was available, that updated and refined them) for the larger components of the game.
Then i use Sonnet for lightweight things, Opus for building P2P systems easily. You just have to prompt it well. 

If these are done, it won't consume as much tokens since it doesn't need to re-read and understand your whole infrastructure on every session.

 

You'd probably get better results with something like https://colbymchenry.github.io/codegraph/

[Matteo 69]

1 hour ago, hvi said:

That sounds like hitting limits to me

You'd probably get better results with something like https://colbymchenry.github.io/codegraph/

After all, it will hit the limits anyway. Its way enough for metin2 development. Its way enough for me.

[eddR 127]

Honestly If I'm using AI it has to be claude max coz other wise it just feels like im talking to a dumbo. Cursor is good for your pythons/ui stuff. Codex sucks dont use it.

Just save up and get max.