I have been spending a lot of time with GPT 5.6 and the new model lineup from OpenAI: Sol, Terra, and Luna. Honestly, everyone is already talking about how these models are beating the competition on benchmarks, and I am sure you have already seen those videos. I am not going to do that again.
What I want to do today is something different. I want to tell you why these models are great. What has OpenAI actually done technically to make these models perform better and at a much lower cost? Because cost was a really big issue before. So whenever the next new model drops and someone tells you it is the best thing ever, you will actually know what is happening inside, not just what the benchmarks say.
What Is GPT 5.6 and Why the New Naming?
The first thing that confused me, and probably confused you too, is the naming. Sol, Terra, Luna. This is not like the usual release, mini, and nano versioning we were used to.
Here is how it works. You have three models: Sol, Terra, and Luna. What is interesting is that unlike older releases where the bigger model had higher reasoning and the smaller one was just faster, here OpenAI is saying that speed is really fast across all three models. Reasoning is best with Sol, less with Terra, and least with Luna. So speed is no longer just a small model advantage.
And pricing is a really big deal this time. This is one of the cheapest model lineups, especially when you compare it to previous releases. In older releases, OpenAI never really talked about cost. They talked about scores and tokens consumed. But now we have open-weight models in the market, models you can actually download and run on your own machine, that are incredibly capable. So OpenAI has had to respond on price. That is why this time around, cost is front and center.
Why GPT 5.6 Is a Game Changer in AI Technology
If you check OpenAI’s own page, they are showing a graph of API cost versus score, comparing GPT 5.6 against models like Agents Lab Exam, Artificial Analyst, Coding Agent, Terminal Bench, and Deep SWE. And GPT 5.6 is winning everywhere, at a fraction of the cost of what Fable 5 used to take. In one example, GPT 5.6 scored 43 on Agents Lab Exam for just $89 in API cost, while Fable 5 needed $1,000 to score less. That is the kind of cost efficiency we are talking about.
How to Access GPT 5.6 and the Confusing Plan Structure
Before we get into the technical stuff, let me quickly explain how you can actually use these models, because it is genuinely confusing.
If you have a free account with ChatGPT, you cannot use these models. By default on a free account, the model you will have access to is GPT 5.5.
For paid users, here is where it gets complicated. On the web interface, in chat mode, you can select your model. But in chat, you do not actually have access to Terra and Luna, even on a paid plan. You get Sol in chat mode with options for medium or high intelligence.
If you go to the Work section, which is like a more advanced version, then you can select all three: Sol, Terra, and Luna. You can also set the effort level: light, medium, high, extra high, max. And you can choose between standard mode and fast mode, which I think used to be called pro mode a couple of days ago. They changed it.
Then there is Codex, which is the separate coding interface. In Codex, you can also select all three models, plus older models like 5.4 and 5.4 mini. And Codex has ultra effort instead of max, on some plans. It is very confusing. I tried to figure out exactly which features are on which plan and honestly gave up. The distribution is inconsistent across web, mobile, and Codex.
The Technical Things That Make GPT 5.6 Better
1. Programmatic Tool Calling
The first big technical upgrade is programmatic tool calling. In simple words, GPT 5.6 can write JavaScript code on its own to call multiple tools, pass results between those tools, and process the final answer, all inside a hosted runtime. You do not need to do anything.
Here is how it works step by step. The model first understands your task using its own reasoning tokens. Then it writes JavaScript code. That code calls Tool A. The result from Tool A gets passed directly into Tool B. Then it processes the final answer from everything.
Before this, if you wanted multiple tools to work together, you had to wait for Tool A to finish before Tool B could even start. Now the model handles all of that automatically. This is a massive win for agent workflows and anything that involves multiple steps or data sources.
2. Multi-Agent Support in GPT 5.6 (Beta)
The second thing is multi-agent support, which is currently in beta. This lets one GPT 5.6 instance coordinate with multiple sub-agents working in parallel and then synthesize all their results together.
This is similar to what they call ultra mode in Codex. When you enable ultra mode, four agents work on your task simultaneously. So instead of one agent doing everything step by step, four agents work in parallel and the whole job gets done much faster. Through the API, you can actually run up to 16 agents in parallel if you need to.
Here is where it gets a little confusing: multi-agent exists as a beta feature on all GPT 5.6 models through the API, but then there is also a dedicated ultra mode in Codex. Why have
both? I am not entirely sure. Through the API I understand, you have more flexibility to configure how many agents run. But having a separate dedicated mode on top of the API feature does add to the confusion.
3. Explicit Prompt Caching
This one is really interesting. Before GPT 5.6, prompt caching was something the model did internally and you had no control over it.
Here is how prompt caching works in general. When you send a prompt, the model checks whether it has already processed something similar. If the beginning of your prompt matches something it has seen before, it reuses that cached processing instead of doing all the work again. This saves tokens and speeds things up.
Before, the model decided on its own when to use the cache. Now, with explicit prompt caching, if you are using the API, you can tell the model directly to cache a specific prompt. You get a cache ID. Next time you send a similar request, you reference that cache ID and the model skips reprocessing everything from the beginning. You consume fewer tokens because you are telling the model exactly what it can reuse. This is a big deal for anyone building products or running high-volume pipelines on the API.
4. Persisted Reasoning
GPT 5.6 can now reuse reasoning items across multiple turns in a conversation. What this means is that the model does not have to start its thinking from scratch every single time you send a new message. It can cache and carry forward the reasoning context from earlier in the conversation, which improves both the quality of multi-turn responses and the efficiency of token usage.
5. Max Reasoning Effort
Alongside the standard effort levels, none, low, medium, high, extra high, GPT 5.6 now supports max reasoning effort for tasks that need the deepest exploration and verification. This is for the hardest problems where you really need the model to be thorough. It consumes more tokens but the quality of the output is noticeably better. One thing I noticed in the documentation is that fast mode consumes 1.5x tokens.
6. Token Efficiency
One of the quieter but important improvements is that GPT 5.6 reaches the same quality with fewer output tokens. Some comparisons suggest it is about one third the token output compared to older approaches to reach similar quality. That directly reduces cost when you are calling the API.
7. Better Front-End Design
This one surprised me. GPT 5.6 creates noticeably better websites and UI than before. I tested this myself. I gave it a prompt about GPT 5.6 and asked it to build a website explaining the model. The result was genuinely beautiful. Better internal understanding of layout, design, and structure seems to have improved with this release.
How the Reasoning Architecture Actually Works
Now let me explain something that ties all of this together: how the reasoning tokens actually work. When you set your effort level from none to max, what you are controlling is how many reasoning tokens the model uses before it gives you an answer.
At low effort, the model responds quickly without spending many reasoning tokens. At max effort, it uses a lot more reasoning tokens to deeply explore the problem, check its work, and verify the answer. Ultra mode adds another layer: multiple agents doing this reasoning in parallel.
This is why effort level really matters. For a simple question, low effort is perfectly fine and much cheaper. For a complex coding task or a research problem, max effort or ultra is worth it.
Why This Matters: The Bigger Picture
The reason OpenAI is making all these moves, cost reduction, explicit caching, multi-agent support, programmatic tool calling, is because the landscape has changed. Open-weight models are now serious competition.
Models like GLM 5.2 are free to download and run on your own hardware. As I mentioned in another video, GLM 5.2 was actually comparable to Claude Opus 4.8 on benchmarks and better than GPT 5.5 in some areas. That is why price has become a central selling point for GPT 5.6 in a way it never was before.
OpenAI still wins on the overall package: the integration, the tooling, the reliability, the multimodal capabilities. But they now have to earn that premium, which is why this release is focused so heavily on cost efficiency alongside raw performance.
Final Thoughts
GPT 5.6 with Sol, Terra, and Luna is genuinely a strong release. The technical improvements, programmatic tool calling, multi-agent coordination, explicit prompt caching, persisted reasoning, and better token efficiency, all work together to make it both more capable and more cost-effective.
But the naming and plan structure is genuinely confusing. Which model is available on which interface, which effort levels exist where, what ultra means in Codex versus what multi-agent means in the API — it takes time to work through all of it.
What I want you to take away from this is simple: next time someone tells you a new model is the best, you now know what to look for. Not just the benchmark score, but the architecture, the caching, the multi-agent design, the token efficiency. That is what separates a marketing release from a genuine technical step forward.
A detailed breakdown of each of these technical points is coming soon. Make sure you are subscribed so you do not miss it.
