What's actually different between GPT-6 Astra, Sol, and Luna?

Okay so here's the lineup, because it trips people up. Astra came out first, a few weeks back, and it's the heavyweight. Most intelligent model OpenAI has shipped, the one you reach for on genuinely hard stuff, deep coding, cybersecurity work, multi-step agent tasks that need real reasoning. It's also the most expensive by a wide margin.

Then this week two more models showed up wearing the same GPT-6 badge. Sol is the middle child, still sharp, still handles multi-step reasoning, but priced for everyday use rather than the hardest one percent of tasks. Luna is the cheap one, built for volume, extracting facts from documents, sorting text, churning out summaries all day without anyone blinking at the bill.

So why not just make Astra better and stop there?

Wait, hold on, why bother splitting the lineup into three at all? If Astra is the best one, why not just make everyone use Astra?

Because almost nobody actually needs Astra for what they're doing most of the time. Think about it. An agent sorting a thousand support tickets doesn't need the model that can architect a cybersecurity system, it needs something fast and cheap that gets the sorting right nine hundred and ninety times out of a thousand. Most real usage, the boring high volume stuff that actually runs a business, doesn't need peak intelligence. It needs good enough, repeated a million times, without the bill exploding.

So Astra stays put as the model for the hard problems, and Sol and Luna exist to hoover up all the volume, everyday work that used to cost more than it needed to.

Here's the twist nobody saw coming

On paper Luna is the budget option, the one you'd expect to lose on every benchmark. It doesn't. On several coding and agent tests, Luna actually beats Sol, the model that costs more. It's not a strict ladder where more money buys more brains, it's a set of models tuned differently for different jobs, and sometimes the cheap one just happens to be better suited to the task in front of it.

Practically, all three are live now inside the usual OpenAI surfaces, and Luna specifically is even available free through the desktop app for lower tier accounts. Which tells you exactly who they're chasing with this release, not the researcher pushing the frontier, but everyone else running the same task a thousand times a day.