One shot, 5.6 Sol on Ultra:
propose a free online dataset that would benefit from a cool 3D animated visualisation and create a deno base webGL 3D animated visualisation to demonstrate it. add some parameter controls suitable to the dataset. allow the user to drag rotate and mouse zoom
Pretty impressive! 
Now see if it can tie it to sun activity and discover something new! Haha. 
(Just saying as I did one like that here somewhere… one of Jay’s threads maybe?
In any case, how long did it take?
All 5.6 models are default "detail": "original". There is no smaller token limit on luna, unlike prior mini, for example. That means if you don’t specify “detail”:“high”, up to 6000 tokens billed.
OpenAI isn’t committing to any promise of the size in billing when you do send a detail parameter, saying instead:
lowandhighcan resize images under their finite limits.
“can” is not a promise. On the API, “high” does seem to be capped at 2500 tokens like others.
For a 640x480 image: I got the same token bill across all these patches models and both Responses and Chat Completions.
Sizing Error: For (1920, 1080), at “high” (not default “original”): which should be Tokens: 2040**, Billed Tokens:** 2448, Chat Completions either has a resize where it is not matching documentation, or a lower bill. On all models.
| model | vision | vision_mult | chat input | calculated | responses input | calculated |
|---|---|---|---|---|---|---|
| gpt-5.6-sol | patch | 1.2 | 2349 | 1951 | 2456 | 2040 |
| gpt-5.6-terra | patch | 1.2 | 2349 | 1951 | 2456 | 2040 |
| gpt-5.6-luna | patch | 1.2 | 2349 | 1951 | 2456 | 2040 |
| gpt-5.5 | patch | 1.2 | 2349 | 1951 | 2456 | 2040 |
| gpt-5.4 | patch | 1.2 | 2349 | 1951 | 2456 | 2040 |
| gpt-5.2 | patch | 1.2 | 2349 | 1951 | 2456 | 2040 |
Calculated column is reversing the NEVER DOCUMENTED after 5.2, 1.2x cost multiplier from the API usage obtained.
Pricing Error: The bill seen on gpt-5.2 means OpenAI is NOT keeping their API contract of maximum 1536 tokens on this model: the call should have been Tokens: 1508, Billed Tokens: 1810 - and not fixed is an even more egregious 5.2 overbilling on large images on that model.
Updated my input image cost calculator before and finalized based on a few verification calls after release - this is based on OpenAI’s documentation, where I have not stress-tested "high" against resolutions, or any kind of operational "low" detail promise that GPT-5 patches models have not had working (despite continuing to offer it in the API shape and documentation):
Plus: models on image/edits fully-included.
25 mins. Probably about 20 confirmations.
Not shabby! I could see it taking someone a few hours or a day at least…
Highly recommended read:
amazing! can u share your prompt bro 
Sure, my pleasure, updated the post
You must check new fields to see how inflated the actual bill is for an API call. You’d discover details reporting that your input usage is a higher cost: always showing almost the entirety of your calls with no cache hits are costing 25% more than the first forum post suggests.
"usage": {
"input_tokens": 1502,
"input_tokens_details": {
"cache_write_tokens": 1499,
"cached_tokens": 0
},
Cache write tokens is an API call being “stored”…for 30 minutes, at higher expense than normally-run input. Yet there is no normal input method to obtain, except for small total tokens <1k.
Meaning: “input_tokens” must not be evaluated directly as input pricing; “write” method is default biling unless there is a cache hit, with a bit more that doesn’t fit the prompts token granularity, apparently with no off switch if you anticipate no reuse, and nothing other than an “explicit” cache for some measure of avoidance, where you set one particular length that can match or miss.
New Responses API usage to collect alongside “usage” - instead of flying blind with unobservable image tool calls powered by user input.
"tool_usage": {
"image_gen": {
"input_tokens": 0,
"input_tokens_details": {
"image_tokens": 0,
"text_tokens": 0
},
"output_tokens": 0,
"output_tokens_details": {
"image_tokens": 0,
"text_tokens": 0
},
"total_tokens": 0
},
"web_search": {
"num_requests": 0
}
},
I hope the infrastructure can cope 
I get the feeling it’s quite a popular release …
I urge the Europeans to get their work done before California wakes up! 
Surely the big story here is just how good Terra and Luna are on higher reasoning levels - why use Sol for agentic coding? It just seems to escalate cost dramatically for little or no gain.
This demonstrates how the smaller models capabilities also appear to scale well with test-time compute - impressive!
Here is why: task difficulty. Real problem solving instead of prompt-to-code, which even gpt-5.2 does well.
In ARC-AGI-2, where running the entire suite costs into the tens of thousands of dollars, this graph format has lower cost at the left, vs the performance height, making the upper-left the area you want to be in unless you must have the hardest problems that can be presented to an AI solved.
There, gpt-5.6-sol dominates the price vs performance and is the standout curve starting at medium (vs the tie of longer reasoning by 3x tokens and higher ultimate cost terra, high)
Had the pro mode have a go at calculations I was doing, and it found two function inputs where Python vs Javascript produced different values, then me needing to discover the truth of the API in undocumented space (Python). 4 runs over two hours; a tad over 100k tests.
Hmmm not convinced. So your argument is that despite great evidence that is it overkill for coding and very expensive for that domain that we should use it anyway because it’s good at solving an advanced puzzle? I’m not sure that’s a good argument but there may be subtleties that will emerge in time that sway the preference.
Not that it’s just good at solving an advanced symbolic reasoning task needing large contextual understanding, such as ARC-AGI-2 with 900 elements to de-pattern - it does that cheaper.
The coding agent index benchmark of the tweet is not merely that of models, it is of the coding harness software on which they run. You as developer cannot directly have questions answered about the model quality you wish to employ when you are not OpenAI selling subscriptions.
Rather that needing to extrapolate out to coding via the opposite “win” case I provide, where the sol model itself is cheaper to deliver the same benchmark score, instead ask rhetorically:
The harness is moot though as consistent for all models?
Well let’s see. I’m firmly in the “Terra Ultra” camp atm and saving trees 
Anyone with a lower end sub is going to be very happy with the utility of the lighter options?
I remain open minded though.
But I’m surprised that I’m even thinking this way now. This shows a major step improvement in smaller model results imho 
Is there any EU date set for Codex ?
Hi @Mb,
It’s already available in the EU, although availability may still be rolling out across accounts.
Thanks for the info; I got it in the chat, but it’s not in Codex yet. So patience is a virtue
True
I live in the EU and I have access to GPT-5.6 (Sol, Terra and Luna) in Codex.
