• This market does not resolve Other, even if the nearest color answer is distant.

@Mira You bought YES for Other; is the quote still correct?

@JacobFriedman Yes. The market will not resolve Other.

@jim I can't accept this color, because I linked to a table of HTML colors and they're RGB not RGBA. I have to calculate a final rendered color using the default theme if necessary.

Has to be a new LLM? Not sure if the recent colour change corresponds to the rollout of the latest GPT-4-turbo or not.

Anyway, the colour has changed, but it's more like it's transparent. So it's a dark grey if you are in the dark theme, and white if you're on the light theme.

Of the existing options, the grey is equidistant from red and blue, and the white is equidistant from yellow, magenta, and cyan.

@chrisjbillington

.text-token-text-primary {
    color: var(--text-primary);
}

.bg-token-main-surface-primary {
    background-color: var(--main-surface-primary);
}

The white is the theme's text color, and the background is the theme's background color(white or black).

3.5 is getting the same color in a chat. So it's not a model-specific color anymore.

So it doesn't have an intrinsic color right now, or the default color using System theme is black which is explicitly excluded.

If OpenAI changes their logo and the color doesn't vary between models, I will calculate the average color of the entire image.

If the theme didn't default to black I would count it and resolve right now to an equal split of every equidistant color. But black is an old color.

You could argue that "the part that varied used to be only the black background, but the part that varies is now the entire logo including a little bit of white. so the average overall color is dark grey." But the only reason it changed was because 3.5 was changed: The latest model would've still been black using the same mask.

So I think I leave the market open, and if it remains attached to the theme, and if the default theme remains black, then it resolves "No new is color is ever issued".

@MiraBot

Orange is the colour they use for "alpha" models (inside /backend-api/models but not in their hardcoded source list). This feature will more than likely continue to be used throughout future releases, so the next official colour will definitely not be orange, or else they'd need to switch this one.

How is “closest” color calculated? Euclidean distance in RGB colors?

Color distance uses the taxicab metric.

I should specify RGB, though it's implied from that being the color specification on the link.

In the current OpenAI logo, the "part that varies" is the circular background and not the white foreground; so if the next foreground is dark it will be ignored.

As in, the dark foreground would be irrelevant to the resolution, or will that entire update be ignored?

Also,
- If the background colour is something more complex, like a dynamic gradient, will it still be resolved with the same algorithm, except perhaps taking the average over a given time interval? But if the average isn't converging, the length of the interval, which seems arbitrary, will determine the average colour. The background colour could also be randomly selected out of a pre-defined pool once per minute, or something harder to test like that.
- What happens if the average colour is exactly half way between the two closest options?
- Does the new model have to be targeted at individuals? The current ChatGPT Enterprise, for example, might have a different colour, but I haven't seen any news articles revealing its UI. The most I saw was a glance in one of OpenAI's demos.

@12c498e

• The "part that varies" determines the area over which the average color is calculated. If the logo is different(color or shape) it won't be disqualified. The dark foreground would be excluded from this area. I could produce a specific mask if necessary, but it's "the green part" of the 3.5 logo. And if they change the logo, hopefully it's obvious what the new "part that varies" is. And if not, it's the entire image(excluding anything transparent).

• For a dynamic gradient, I would first determine the "part that varies" and then take the average over both time and area. The time is until it repeats. If it never repeats, then I'll take 5 minutes of video.

• If it's randomly selected out of a pre-defined pool, I'll resolve to an equal split of every color it could be.

• If it's exactly halfway between two options, I'll resolve 50% to each.

• I think it can be enterprise too. But note that the "speaker logo" is different from the "GPT store logo" or the "web UI logo". I'll need to see evidence of a ChatGPT-style conversation where the speaker logo is a color. Ideally 2 conversations so I can confirm that it's consistent between them.

• The product does need to be announced by OpenAI, so leaked test colors don't count. But ChatGPT Enterprise has already been announced and would count.

• If they get rid of the speaker logo concept in a new ChatGPT-style product, it doesn't count at all so it'll have to wait until it comes to a product that does have speaker logos.

• I expect it to be a standard color or obvious from inspection, but these are specified in case they release something weird.

I've had Orange show up as what looks like a frontend glitch in the past:

https://archive.area17.com/directory/2023_openai/36_brand-guidelines.pdf suggests that OpenAI tends to use these colors:

• White and black as primary colors (though black is disqualified, and white would just look ugly if the logo were to be inverted).

• Red, green (but this is disqualified as already used), and blue as secondary branding colors. Both red and blue are plausible IMHO.

• Yellow, magenta, and cyan are tertiary colors but not that plausible I think?

• Orange, violet, and brown are "supplemental" colors so very, very unlikely I think.

Bet on the next OpenAI GPT logo. This purple has to go!