Eighty percent. That’s how much OpenAI cut the price of GPT‑5.6 Luna on July 30, 2026. It’s the single most consequential number in this entire story, and it has almost nothing to do with the thing people keep asking me about.
Because what lands in my inbox isn’t “should we move our summarization pipeline to Luna now that it costs a fifth of what it did.” It’s “when is GPT‑6 Sol dropping.” Different question. Worse question.
What actually exists right now
As of 2026, GPT‑6 Sol and GPT‑6 Luna have not been officially announced. The current models are GPT‑5.6 Sol and GPT‑5.6 Luna. There is no official information about future GPT‑6 versions. That’s the whole factual picture. Everything beyond that line is somebody’s traffic strategy.
I want to be precise about this, because the SEO ecosystem around unannounced models has gotten genuinely impressive in its confidence. There are pages ranking right now with titles like “GPT‑6 Sol Release Date” and “GPT‑6 Luna Rollout Status for API Teams.” Some of them are honest about it. One of them, to its credit, explains that it checked OpenAI’s model list, the pricing page, the API changelog on developers.openai.com, and the openai.com news feed, and the answer was no, OpenAI has not announced GPT‑6 Sol. That’s actual reporting. That’s a site doing the boring work and publishing the boring result.
Others are less careful. There’s coverage framed as OpenAI “rolling out more affordable GPT‑6 Sol and Luna models,” sitting in a feed next to unrelated items about Apple fitness trackers and Anthropic’s Claude Opus 5.5. You can feel the content mill humming. A version number gets pulled from a rumor, upgraded to a headline, and then quoted by the next post down the chain as though it were a press release.
Why the rumor is the wrong thing to plan around
Here’s what bugs me as a reviewer. Waiting for an unannounced model is a way of avoiding the harder engineering questions about the model you can actually call today.
One of those EvoLink pages made a point I keep thinking about, and it’s the most useful sentence in the entire pile: a low token price cannot tell you whether retries will consume the saving, and a fast single response cannot tell you whether your queue meets its deadline at normal concurrency. That’s the right instinct. Launch news is a trigger to go measure something, not a conclusion.
Apply it to that 80% Luna price cut. On paper, your inference bill drops to a fifth. In practice, your bill is price multiplied by tokens multiplied by attempts. If cheaper capacity means you start retrying more, or you get sloppy about prompt length because tokens feel free, the savings evaporate and you won’t notice until the invoice arrives. GPT‑5.6 Terra got a 20% cut in the same update, which is a much less exciting number and possibly a much more predictable one.
None of that analysis requires a GPT‑6 anything. All of it requires you to instrument your own pipeline, which is less fun than refreshing a rumor page.
The slider is the more interesting story
The GPT‑5.6 Sol update in ChatGPT shipped a thinking slider for Plus and Pro users on web, mobile, and desktop. You choose how much thought the model puts into an answer. Quick for everyday questions, higher when you want it to work harder.
I like this more than I expected to, and not because it’s clever UX. I like it because it makes a tradeoff visible that used to be hidden inside routing logic you had no control over. Latency, cost, and answer quality have always been in tension. Handing users a dial is an admission that no single default serves everyone, which is more honest than pretending one exists.
It’s also the kind of change that shows where products are actually heading. Not a bigger version number. A control surface.
My advice, unglamorous as it is
If you’re building on these models, treat unannounced versions as exactly zero percent of your roadmap. You cannot price something that has no pricing page. You cannot test something with no endpoint. You cannot plan a migration to a product that no company has confirmed is being built.
What you can do: benchmark GPT‑5.6 Luna against whatever you’re running now at your real concurrency, with your real retry policy, on your real prompts. Watch what the thinking slider tells you about which of your tasks genuinely need more compute and which ones you’ve been overpaying for out of habit. Build your evaluation use so that when a new model does show up, you can measure it in an afternoon instead of arguing about it for a quarter.
That’s the unsexy version of staying ahead. It works whether GPT‑6 Sol arrives next month, next year, or never under that name at all.
🕒 Published: