what language model version is used ? is it chatgpt 5, 5.1, 5.4 or what is it ?

[PurpleGreenW]

why do you make changes towards LESS transparency and LESS control.

[mark-at-pieces]

Hi @PurpleGreenW,

Happy to answer this directly.

We recently moved the Copilot from a one-shot prompt system over to a full agentic engine — meaning the Copilot can now plan, take multiple steps, and reason across them rather than just sending a single prompt and returning the result. That architectural shift is actually one of the main reasons we also overhauled how model selection works, since agentic workflows lean much harder on the underlying model's reasoning quality and consistency than one-shot prompting did.

The way model selection works now is that you first pick a provider (GPT, Claude, Gemini, etc.), and then within each provider you pick a tier:

• Fast — optimized for speed
• Balanced — the everyday default
• Extra-thinking — for harder, more reasoning-heavy questions

The same three tiers exist across all providers, but the underlying model powering each tier depends on which provider you picked. On the GPT side right now, that's GPT-5.4 nano (Fast), GPT-5.4 mini (Balanced), and GPT-5.5 (Extra-thinking). Claude and Gemini have their own best-in-class models mapped to each of those tiers.

These tiers are dynamic — as newer models come out and we validate them, the underlying model behind each tier gets upgraded automatically, so you're always on something current without having to track releases yourself.

As for why we moved away from the older granular per-model picker: it came down to response quality, especially under the new agentic engine. With the old picker, it was easy to end up routing to a legacy model that produced noticeably worse answers in the Copilot — and that gets even more pronounced in an agentic flow, where a weaker model compounds errors across multiple steps. Curating into tiers lets us keep a quality floor under every response.

On the broader direction: we're intentionally making model selection less technical for the average user — pick a provider, pick a tier, done. That said, granular per-model selection as an advanced setting is on the roadmap for users who want that level of control. Not shipped yet, but it's planned.

Thanks for the feedback.

[szabolcs-dekany]

So, the fact that you are saying you are using GPT.5.5 is just not true. I just got an error from openRouter (the request timed out) that said the model was openai/gpt-5.2-20251211 I had extra thinking with reflection on. Care to elaborate? I think thats the answer @PurpleGreenW. Very disappointed where this is going. Sadly cancelled my sub after this. Not worth it.

Not gonna share the whole error JSON but here's proof:

{
  "type": "response.failed",
  "response": {
    "id": "<not-gonna-share-this>",
    "object": "response",
    "created_at": <not-gonna-share-this>,
    "model": "openai/gpt-5.2-20251211",
    "status": "failed",
    "completed_at": <not-gonna-share-this>,
    "output": [],
    "error": {
      "code": "server_error",
      "message": "internal stream ended unexpectedly"
    }
  }
}