GPT 5.6 Just Dropped: Sol, Terra & Luna Explained
You will know what GPT 5.6 Sol, Terra and Luna are, what the early benchmarks and pricing suggest, and how to get ready to test them in Codex when they reach you.
O que você precisa
A ChatGPT account (for Codex access)
Optional: a Claude account, to compare against Claude Code
A healthy dose of skepticism until you can test the models yourself
Passo a passo
Know Sol, the new flagship
Treat Sol as the biggest of the three GPT 5.6 models and the new flagship. The benchmark chart also shows a Sol Ultra version. At the time of recording it was announced about an hour earlier, so nobody had tested it yet.
Look at Terra for everyday work
Consider Terra, the middle ground: it is announced to give GPT 5.5 level performance at half the cost. I rate GPT 5.5 as one of the best models for coding, following instructions and business use cases, so Terra is the one to watch if cost matters.
Treat Luna as the lightweight option
Place Luna as the smallest model. Its capability is unclear, maybe around GPT 5.4. It is also unknown whether any of the three models use a new architecture or the previous one.
Expect a limited rollout
Do not expect immediate access. Because of cybersecurity concerns around frontier models (the Fable 5 situation), only a small group of people gets GPT 5.6 first, before a wider rollout.
Read the Terminal Bench results with care
Check the Terminal Bench chart: both 5.6 Sol and Sol Ultra score slightly above Fable 5 and Mythos 5, roughly 4% to 8% over Fable 5. Token usage also goes up with GPT 5.6. Remember these are launch numbers, not your own tests.
Compare the pricing
Compare output prices: I believe Fable 5 costs about $50 to $75 per one million output tokens, and standard Sol comes in at nearly half of that or less. For value per dollar, OpenAI stays ahead.
Use Codex for coding on a tight budget
Open Codex for coding work, especially if your budget is tight: I get much more usage out of it than out of Claude Code, where I run out of usage when I use it a lot. I still run both every day.
Codexhttps://chatgpt.com/codex
Claudehttps://claude.ai
Keep Claude for front-end design for now
Be careful with GPT models on front-end design (the look of the app, buttons, layout): I find them bad at it, though this version is said to improve. Fable 5 currently handles both UI and back end well, so test before you switch.
Stack Codex referral codes
Collect Codex referral codes now. Each one resets your token usage and stays valid for a month, so when GPT 5.6 arrives you can reset your usage several times and test the model fully.
Fique atento a
Every new model looks like the best one on launch day: stay skeptical until you have tested it yourself.
GPT 5.6 uses more tokens than earlier models, so a lower price per token does not guarantee a lower bill.
GPT models have been weak at front-end design, so do not move UI work over without checking the results.
Access is limited at first, so do not plan work that depends on GPT 5.6 until it reaches your account.
