01What is GPT-5.6 and how it works
GPT-5.6 is OpenAI's latest model release, and it comes as three different sizes: Sol, Terra, and Luna. They all share the same architecture but are built for different needs, from heavy-duty coding to simple everyday tasks.
Sol is the largest and most capable, Terra is the balanced everyday model, and Luna is the fast and cheap option. The models have a 1 million token context window and can output up to 128,000 tokens. They were delayed by a government review but are now generally available, rolling out slowly worldwide.
In this episode, Julian Goldie tested all three models on his GoldieBench leaderboard, building 33 different projects with them. He found that Sol could create games, videos, landing pages, and more, all from a single prompt, with impressive quality.
02Main features and what's new
The big news is that GPT-5.6 is not one model but three, each with distinct strengths. Here are the key features:
- Sol: The flagship model, designed for complex, multi-step agentic work. It's the most expensive but delivers the best results.
- Terra: The everyday production model, balancing cost and performance for typical tasks.
- Luna: A fast, cheap model for high-volume, latency-sensitive work. It costs 80% less than Sol.
All three models share a 1 million context window and a 128K max output, with a February 2026 knowledge cutoff. They can be used via OAuth, meaning you don't need an API key, which can save money if you have access to platforms like Codex or Hermes.
Julian highlighted that GPT-5.6 supports a 'goal mode' in Codex, allowing agents to run for hours or days until a task is done, which is a powerful feature for autonomous work.
03How to set up and get started with GPT-5.6
To start using GPT-5.6, follow these steps:
- Wait for the rollout to reach your region, as it's being released gradually.
- Choose the model that fits your task: Sol for heavy lifting, Terra for daily use, Luna for volume work.
- Access GPT-5.6 through platforms like Codex or Hermes, preferably with OAuth to avoid API costs.
- If using Codex, switch between models in the dropdown, or set up a separate profile in Hermes.
- Use the goal mode in Codex for long-running autonomous tasks, telling the agent to not stop until the job is done.
- Try building a small project, like a landing page or a game, to see the output quality.
Julian recommends avoiding ChatGPT for serious work; instead, use an agent operating system like AgentOS to get the most out of GPT-5.6.
04Using GPT-5.6 for games, videos, and more
In the episode, Julian showed several examples of what GPT-5.6 can create. Sol made a game that looks like Fable 5 in quality, an open-world game that felt smooth and fun, and a Doom-style game that ran perfectly. It even created a full video animation with a progress bar, all from one prompt.
He also built landing pages and an analytics dashboard. The landing page from Sol looked professional, with nice colors and a clean UI, unlike typical AI slop. Terra and Luna also did decent jobs on landing pages, but Sol's output was noticeably better.
When comparing GPT-5.6 Sol to Fable 5, Julian found the results very similar in many cases. For an arcade game, Sol felt slightly smoother, and for a 3D space game, Sol absolutely crushed it—Fable 5 couldn't even turn the camera around, while Sol was smooth and easy to use.
05Results, tests, and comparisons from the episode
Julian ran tests on GoldieBench, building 33 projects. He gave each model the same task—a 3D space game—with one shot. Sol took 6 minutes and cost $1.18, Terra took longer but cost $0.43, and Luna was fastest at 1 minute 37 seconds and cost only $0.12.
All ran error-free.
Token output also varied: Sol produced more code but at a slower speed of 109 tokens per second, while Luna hit 198 tokens per second. On benchmarks, Sol scored 59 on the intelligence index, just one point below Fable 5's 60. On Terminal Bench 2.1, Sol scored 88.8, Sol Ultra 91.9, and Fable 5 83.4.
Julian noted that Sol's creations were far better than Terra's or Luna's—'not even comparable.' However, he warned not to trust benchmarks blindly; testing yourself is key, which is why he built GoldieBench.
06Tips, limits, and mistakes to avoid
One big tip is to switch models based on task difficulty. Use Luna for high-volume work, Terra for daily tasks, and Sol only for the hard 20%—complex coding or creative projects. This keeps costs down without sacrificing quality.
Another tip is to use OAuth to avoid API fees. Julian used it to build everything for free, saving money compared to models like Fable 5, which require payment in agents like Hermes.
Be aware of limits: some UI elements, like buttons in the dashboard, may not work. Also, Sol is slower and more expensive, so don't use it for simple tasks. Finally, don't just use ChatGPT—use an agent OS to get full benefits, but if choosing between Fable 5 and Sol, Julian personally prefers Fable 5.
07Who GPT-5.6 is for and why it matters
GPT-5.6 is for anyone who builds with AI, from developers to creators. Sol is perfect for complex projects like games and videos, while Terra and Luna suit everyday use and high-volume tasks.
This release matters because it offers flexibility and cost savings. You can choose the model that fits your budget and task, rather than paying premium prices for everything. Julian believes this is the best release from OpenAI this year, aside from their image model, and it's a strong competitor to Fable 5.
If you're serious about AI, using GPT-5.6 within an agent operating system can multiply your output, as shown in the episode. The goal mode in Codex also enables long-running autonomous work, which is powerful for automation.
