gpt-5-Codex is available on openrouter

Stream: Machine Learning for Theorem Proving

[Deleted User 968128]

https://openrouter.ai/chat?models=openai/gpt-5-codex

For those who don't use openrouter, I encourage you to take a look. No subscription required as everything is pay as you go. There was a 50% discount on gpt5 last week, which was quite nice. Also, new and highly capable models are being released weekly at this point and it has a very good chat interface with pinning.

Just released QwenMax (scores 100% on AIME2025 with tools) is also now on openrouter and I'm curious how it will perform. https://openrouter.ai/qwen/qwen3-max

In order to keep costs down, I will frequently switch to a cheaper model for certain prompts, many of which are free and don't consume as much power/water per prompt (I consume about one 5m shower per day in AI usage on average). Do be warned however, the free models do not have Zero Data Retention (ZDR) and are training on your prompts (which they state clearly).

My monthly spend ends up being quite low on openrouter, even with very heavy agentic usage - one of the reasons I dropped the openai pro account for $200 monthly. And I also get the benefit of trying out every model as it is released.

requesty.ai is a competitor that seems to offer a comparable service, but I do not use them.

[Gavin Zhao]

Just curious, what coding agent you're using? Are you just using codex and overriding the base_url to OpenRouter?

[(deleted)]

Also what have you proved

[(deleted)]

I find that GPT-5 loves deleting the whole code when it feels hopeless. So GPT-5 can't run without supervision

[Deleted User 968128]

Gavin Zhao said:

Just curious, what coding agent you're using? Are you just using codex and overriding the base_url to OpenRouter?

OpenAI built clients are built for the OpenAI apis. Afaik, this is not recommended.

[Deleted User 968128]

In general, the auto coding agents are only useful for specific tasks and not yet broadly applicable.

[Deleted User 968128]

Not much, like most, I agree ML is at a fairly early stage. Auto coding is more advanced, but also quite problematic.

[Deleted User 968128]

One thing I would strongly recommend is not to use auto coding if you are trying to learn something, like RL. I'd even call that an anti-pattern. Auto Coding tools are not meant for that.

[(deleted)]

And I know that janitorial work is quite valuable. Fixing AI mistakes is more important than I thought. Because AI is quite dumb.

[(deleted)]

I spent time cleaning up proofs, and turns out the time spent to clean up is valuable. Not a trivial step that AI can easily do, for now.

[Deleted User 968128]

I think they'd be very useful for bumping versions. They can handle anything mechanical very very well.

[(deleted)]

Now that I think of it, to observers, my attitude towards AI has always been quite contradictory. I complain about AI probably even more often than praise AI

[(deleted)]

But whatever. It is a productivity boost. I can quantify it in terms of money, it doubles my income. Even with all the rough edges

[Deleted User 968128]

AI is at the junior dev level where you know they will be promising after some coaching, but right now are a bit of a hindrance.

It is self evident they are on the fast track for promotion. Performance reviews have made clear however they need to start focusing more on the important stuff.

eg: less voice tech, more lean