It will use a new architecture meaningfully different from GPT-4

"GPT-5" refers to a model named or presented as "GPT-5"(or a variation).

It will use a new architecture meaningfully different from GPT-4

https://openai.com/index/introducing-gpt-5/

GPT‑5 is a unified system with a smart, efficient model that answers most questions, a deeper reasoning model (GPT‑5 thinking) for harder problems, and a real‑time router that quickly decides which to use based on conversation type, complexity, tool needs, and your explicit intent (for example, if you say “think hard about this” in the prompt). The router is continuously trained on real signals, including when users switch models, preference rates for responses, and measured correctness, improving over time. Once usage limits are reached, a mini version of each model handles remaining queries. In the near future, we plan to integrate these capabilities into a single model.

I'm not sure I should count this. You can select 4 levels of reasoning - that might skip the router.

Google Summary says it's a "256 expert MoE with 8 experts active per token" but it's citing some guy on Youtube. I've heard other rumors that it's a continuation of the 4o base model.

I did write:

Benchmarks and features are not considered: Only how the model is presented by OpenAI.

So since OpenAI is presenting it differently, I guess I'll resolve YES

OpenAI will claim it faster than GPT-4 Turbo

GPT-4 Turbo I see non-OpenAI claims of 20 tokens per second. GPT-5 is between 38-100 tokens/second. So it seems to be faster. But strictly speaking I need a claim from someone at OpenAI that one measured number is larger than another...

The model comparison page shows 3 speed symbols for both:

It is faster, but I need a chart or reported metrics from an OpenAI tool or site to resolve YES. Otherwise, it's NO. I'll give you guys some time to find charts.

@creator now that you're back, maybe we can resolve this? I'm not sure how

@mods It's been over 3 months, can. you resolve the "up to 3 months after release" questions?

@jim @mods resolves NO

@mods Can you guys resolve this since Mira deleted her account

@Vesperstelo Any specific ones you'd like resolved still? Suggested resolutions and citations?

it appears that the different GPT-5 models have various different knowledge cutoffs. The ChatGPT versions (thinking and non-thinking) report “June 2024”. But the API documentation says “May 30th, 2024”, “September 29, 2024”, or “September 30th, 2024” depending on the model version.

https://platform.openai.com/docs/models/compare

this system prompt from Pliny says June 2024 https://github.com/elder-plinius/CL4R1T4S/blob/main/OPENAI/ChatGPT5-08-07-2025.mkd

@BionicD0LPH1N

I'm not sure how this should resolve. When using it via ChatGPT, it's a model router, but when you're using it via the API, you're picking a model directly.

@SimonWestlake This reads to me like the models themselves (what you can access through the API) don't have a built-in router, so it's not a core part of GPT-5 itself. The gpt-5-chat model in ChatGPT UI is a separate model that analyzes the query and makes a route call to the appropriate version of GPT-5, but this wrapper model itself is not part of the release and cannot be queried directly (since it's a pass-through layer). So GPT-5 is packaged with a router, but it's not a router itself, and the router is likely an extension of their input safety layer which had always been a separate model.

@moozooh Yeah I think that's the most reasonable interpretation.

@SimonWestlake I am of the opinion that it should resolve YES. This is roughly what I had in mind when I created this option.

See /Bayesian/will-gpt5-involve-a-model-router

@VitorBosshard Disregard my original comment, I think this should resolve NO

@Mira How will questions like this resove if GPT-5 has configuration options that allow it to be on either side of the line?

@jim resolves YES if a one-sentence, plain-English prompt is found that allows GPT-5 to solve jim's agents benchmark at least 20% of the time.

@jim This is a wild benchmark. I have some level of general intelligence and there is no way I could do this.

@MalachiteEagle I don't expect "open"AI to disclose such information to be honest. They don't even publish parameter couts anymore let alone training data details.

@ProjectVictory maybe. Guess we'll find out. Think they like to sell a narrative around the models they release, that narrative might include something like this.

@mods And this one resolves NO.

@mods Resolves YES.

@mods Resolves NO, along with the option for May - October.

The new "GPT-4o" counts as GPT-4.5 in the paired market, since it was based on the name. "A model with separate branding in the GPT-4 series". So I've resolved the 4.5 option here to match.

"GPT-5" designation here is also based on the name. I will not be looking at benchmarks, only how the model is presented by OpenAI.

@Mira wait if you consider gpt4o as gpt4.5, will the model Jimmy calles gpt4.5 here resolve as gpt5? https://x.com/apples_jimmy/status/1833595024543781088

@Phill It has to be named GPT-5. Or presented as the "fifth generation of the GPT series". So that name is unlikely to count.