GPT 5.6 Just Dropped: Sol, Terra & Luna Explained

OpenAI just dropped GPT 5.6, literally an hour before I recorded this, and it arrives as three new models: Sol, Terra, and Luna. In this breakdown I go through what each model is, the politics around the limited rollout, the early Terminal Bench numbers against Fable 5 and Mythos 5, the pricing that genuinely shocked me, and one practical tip to squeeze the most out of Codex when these models land.

1. Meet the lineup: Sol, Terra, and Luna (0:00)

Sol is the biggest of the three and becomes the new flagship model. Terra is the middle ground, and it is the one that looks extremely interesting on paper: it gives you access to GPT 5.5 level performance at half the cost, which is unbelievable. Remember what GPT 5.5 is: one of the best models for coding, one of the best at following instructions, extremely good at business use cases, and quite affordable for the performance it delivers. Getting that level at half price would be a serious deal.

Luna is the lightweight option, and honestly we do not really know where it stands in capability yet, maybe around GPT 5.4. One more open question across the whole family: we do not know whether these models are a new architecture or the same architecture they had before. OpenAI promises a big level of improvement across the board, as every launch does.

2. Limited rollout and the Fable 5 situation (1:22)

The release arrives under unusual conditions. GPT 5.6 has taken a hit in its rollout because of the Fable 5 situation. AI is getting very political right now, because frontier models have become so good that a simple GPT or Claude subscription could carry major cyberattack and cybersecurity risk, which is absolutely insane to say out loud. The practical consequence: only a small group of people gets access to these models first, before they roll out across the board.

3. What happens to open source AI (2:02)

That gating raises a question I covered in a previous video: what happens with open source? GLM 5.2 is getting near the level of these frontier models. What really happens when open source reaches that level and those open models become available to everyone with a powerful enough computer? I do not have the answer, but it is the trend to watch while the frontier labs restrict access.

4. Benchmarks: Terminal Bench (2:35)

Every provider strives to look best on launch day, so let us look at the numbers with that in mind. On Terminal Bench, one of the staple benchmarks, there are actually two versions of the flagship: Sol Ultra and 5.6 Sol. Both outperform Fable 5 and Mythos 5. For context, Fable 5 is the smaller, more guardrailed version of Mythos 5, and both new Sol variants sit slightly above it. Against Fable 5 that reads as an increase of roughly four to maybe eight percent in performance. Token usage also goes up in GPT 5.6, which is the standard thing we see across the board with these releases. We still do not have access to the model, so treat all of this as early signal, not verdict.

5. The pricing shock (4:08)

Here is where I was genuinely shocked. Sol, the standard Sol, positioned a level above Mythos 5, comes in at a much more affordable price. Fable 5 sits at, I believe, fifty to seventy-five dollars per one million output tokens. Sol is nearly half of that, if not more than half off. OpenAI is still leading the race on value per dollar. They have access to a lot more computing power than Anthropic, which lets them price more aggressively. Whether their architecture is also more efficient, we do not know, but for everyday usage this is going to be a serious contender.

6. Coding: GPT vs Claude, and why I say use Codex (5:06)

My honest position: I love Claude and I run it every single day, both Claude Code and Codex side by side. But I do believe GPT 5.5 is a better coding tool than Opus 5.8. I cannot compare it to Fable 5 yet because I have not tested that matchup, and Fable 5 really was fantastic. What I know from daily use is that when I lean hard on Claude Code, I simply run out of usage. So right now I highly recommend using Codex more than Claude Code, especially if you are on a tight budget: your capacity to use Codex is largely superior, you just get more value out of it.

7. The front end weakness (5:34)

One personal note that could decide the next round: GPT models have been very bad at front-end design, simply terrible. But I have seen lately that they are improving this in this version of the GPT models. If OpenAI gets front-end quality, the actual look of a web application, the buttons, the design, to match its raw back-end code strength, it will be very hard for Anthropic to keep up. Right now Fable 5 is capable of extremely nice UI and back-end both. If Sol reaches that level too, things get very interesting.

8. Tip: stack Codex referral codes (6:25)

Practical move you can make today: Codex is giving out referral codes that reset your token usage, and they are valid for a month. Pile them up now. When GPT 5.6 comes out, you can reset your usage time and time again and really milk everything you can from this model.

Pitfalls and tips

• Every model is "the best one" on launch day. Stay a bit skeptical until you can run it on your own work.
• Do not read per-benchmark wins as a full picture; we do not even know yet if this is a new architecture.
• If budget matters, route heavy work through Codex and save Claude usage for where it shines.

Where to go next

For the other side of this model race, read my verdict on Anthropic's latest mid-tier release in Is Claude Sonnet 5 Actually Worth It?.