Introducing GPT-5.6 series: Sol, Terra and Luna. Coming July 9 10am PT

Announcements
27

What is expected from past ChatGPT account behavior: Purchased Codex credits, at the large dollar increments offered in ChatGPT, are billed and consumed at the same ultimate dollar rate as the API published token billing. Those somewhat obfuscated “credits”, not valued directly in dollars or API fun bucks, thus can be translated with a bit of thought currently - and with gpt-5.6 solar bodies here showing the same pricing as gpt-5.4 and gpt-5.5.

Not clarified for you: the exact dollar amount included amount in a “reset”, or included in the shorter quota period, or generally included in your subscription. You probably get a deal there as a frequent user, all things considered.

sps 28

UPDATE: GPT-5.6 Sol, along with Terra and Luna, will launch publicly on Thursday, July 9.

13 Likes
30

Oh my! I can’t wait for Thursday to arrive!
Go time, go!

For anyone else who is on the fence like me:
What’s the first thing you are going to do with the new models?

31

Do you have your timer set?

I am too so excited and I actually have been checking if by any chance I would have gotten an early access…no such luck​:sweat_smile: The first thing I am going to do is to test my new persona, that I am so excited to test.

32

Wow. That must be quite the complex personality if it needs that much reasoning to get the expected behavior.

I’ll have it work on my favorite side project: rebuilding the navigation layer from one of my favorite old-school video games. Then it can build the tools for difficulty balancing and story interactions.

I should probably spend the rest of the day preparing the context and prompts instead of writing here.

33

Hah it’s not “be a funny assistant” kind of persona…Believe it or not, this one actually is complex because it is not just fixated on one personality. This SKILL will depend a lot on reasoning and matching the users input and following the SKILL, so that persona holds even when stress-testing it.

Oh that sounds interesting and fun!

34

Benchmark them against the other models we use for coding tasks and slot them in where they fit.

35

“Launch publicly” does not state who, where, how.

There is no API documentation change for the methods discussed in the model blog (such as explicit cache storage parameters at 1.25x cost), no disclosed updated information such as vision technology and its billing, and API SDKs are stalled at “two weeks ago”. Not even a background picture for the model on the platform site.

Completely embargoing implementation details up to now of what has been announced on social media and what is being employed in alpha and in preview (for on-switch ready API jump-started partners) seems silly, only serving to deepen the timeline handicap between copacetic partners and not.

So, predict: “for ChatGPT Pro 20x users”?, or “Only via Codex app?”. When ID verified? When you complete this government-disclosed form?

UPDATE: answer: the model is all over the place with no noted restrictions, rolling out over 24hrs.

36

Any news on when it will be released?

37

Mr Bean Waiting Impatiently in a Field
Waiting, Waiting, Waiting…

38

See discord and their announcement times for previous releases
Basically always around 10AM PDT (up to 3 hours later, 1 time half hour before)
Most announcements/releases happen around that time

sps 39

Today. 10am PT.

3 Likes
40

Superapp with super models?
Let’s go!

41

@here

Join the livestream.

https://openai.com/live/

42

Interesting that the new model announcement is a sideshow to product announcements.

43

Extracted API notes

  • Models: gpt-5.6 → gpt-5.6-sol; also gpt-5.6-terra for cost/perf, gpt-5.6-luna for high-volume efficiency.

  • Reasoning effort: none | low | medium | high | xhigh | max; benchmark current setting and one level lower.

  • Pro mode: use reasoning.mode: "pro"; not a separate model slug. reasoning.effort stays independent. Higher token count, same rate.

  • Persisted reasoning: set reasoning.context to auto, all_turns, or current_turn.

  • Prompt cache: implicit still works; explicit mode via cache breakpoints / prompt_cache_options.mode: "explicit".

  • Cache TTL: replace prompt_cache_retention with prompt_cache_options.ttl.

  • Cache pricing: cache writes cost 1.25× uncached input; monitor cached_tokens and cache_write_tokens.

  • Programmatic Tool Calling: add programmatic_tool_calling; opt tools in with allowed_callers; handle program, function calls, and program_output.

  • Multi-agent beta: Responses API beta for parallel subagents and synthesis on separable complex tasks.

No “snapshot date” in model names.

44

Looks like the token price of Sol is identical to 5.5.

It’s nice to see a lack of an increase. :tada::rose:

It’s also worth noting I see all the new models in the latest version of Codex CLI (0.144.0), nice! I had expected not to see them as based in UK and feared they might be a US exclusive.

And just look at the reasoning options now!:+1:

image
image1202×310 66.9 KB

Ultra looks particularly interesting …

45

IMG_4628
IMG_46281170×2532 168 KB

Well now I have finally access to 5.6 Sol​:tada:

46

Yay! All the new models work in Norway!