GPT-6 Sol and Luna: What OpenAI's New Cheap Tiers Actually Change

On September 22, OpenAI released two more members of the GPT-6 family: Sol and Luna. Both are cheaper than what came before, and both are described as bringing the same gains as GPT-6 Astra down to a lower price. That headline is easy to read past, and the part underneath it is the one worth stopping for: “Sol” does not mean what it meant two months ago.

If your team standardized on a tier by name rather than by testing, this release is the reminder to check that the name still points at the same thing.

The Tier Names Moved Again

OpenAI’s GPT-5.6 family, released in July, put Sol at the top as the flagship, with Terra as the mid tier and Luna as the fast, cheap option. GPT-6 changed the hierarchy. GPT-6 Astra shipped first, on September 3, as the new flagship, aimed at agentic and computer-use tasks and gated to partners in an application-based program. Sol and Luna arrived nineteen days later as the two tiers underneath it, not above it.

So the same two names now sit one rung lower in the lineup than they did in the previous generation. A person who picked “Sol” in July because it was the best model available is not automatically using the best model available today, and a person who assumes today’s Sol is a direct upgrade of July’s Sol is assuming something OpenAI never claimed. Why AI model version numbers tell you nothing covers this pattern across labs: names and numbers describe a position in a lineup, not a fixed capability level, and the lineup itself gets reshuffled without warning.

What Sol and Luna Actually Cost

The prices are concrete, at least. OpenAI lists GPT-6 Sol at $2 per million input tokens and $10 per million output tokens, and GPT-6 Luna at $0.10 and $0.50. Cached input reads carry roughly a 90 percent discount on both tiers. Coverage of the release puts the cut at close to half of the equivalent GPT-5.6 pricing, continuing the trend this blog tracked with OpenAI’s August price cuts.

The two tiers are positioned differently, not just priced differently. OpenAI frames Sol for complex work done repeatedly, the kind a knowledge worker or developer does many times a day, and Luna for high-volume, narrowly defined jobs such as summarizing, extracting, or answering a straightforward question. For a person rewriting work messages, that maps loosely onto “drafting something that needs judgment” versus “cleaning up something short and simple”, though the boundary is a marketing description, not a line you can verify without testing your own material against it.

“Professional Work and Factuality” Is a Claim, Not a Benchmark

OpenAI’s own announcement frames it as Sol and Luna being trained with methods similar to Astra, carrying forward gains in professional work, factuality, coding, computer use, and alignment at a fraction of Astra’s cost. That is a real claim, and it may hold up. It is also, as of this writing, a company’s description of its own release rather than an independent comparison on ordinary rewriting tasks.

That gap matters because “professional work” and “factuality” are exactly the phrases a benchmark struggles to capture. Do frontier models write better emails found that the largest, most benchmark-decorated models do not reliably beat much smaller ones at rewriting a three-paragraph message, because rewriting is a narrow, constrained task rather than an open reasoning problem. A model can post real gains on coding and agentic benchmarks and show a small or unclear improvement on “make this firmer without sounding rude.” Nothing in the September 22 release changes that argument. It adds one more model that has not yet been tested against it.

What to Check Before You Assume Sol Beats Old Sol

The mistake this release invites is treating a familiar name as a stable fact rather than a label that moved.

Before:

Just use Sol for this, it’s the good one, we tested it back in July.

After:

Test whichever GPT-6 tier is current before relying on it for client work. OpenAI renamed the lineup between the 5.6 and 6 generations, so “Sol” today is not the model we tested in July.

The second version survives the fact that a name got reused for a different position in the lineup. The first one quietly assumes stability that the release itself contradicts.

A Wrivio Context for evaluating a new model release could say:

Summarize this announcement as three plain facts: what changed, what it costs, and what is still an unverified claim by the vendor. Keep every number and date exactly as written. Do not add a recommendation.

Press Ctrl+Shift+Space, paste the announcement or article, and check the diff against the original. A summary that keeps the numbers intact and drops the marketing framing is the one worth acting on.

Common Questions

Is GPT-6 Sol the same model as GPT-5.6 Sol?

No. The name carried over but the position in the lineup did not. GPT-5.6 Sol was OpenAI’s flagship at the time; GPT-6 Sol is a mid tier underneath the newer flagship, GPT-6 Astra.

How much cheaper is GPT-6 Sol and Luna than the previous generation?

Reporting on the release puts the cut at roughly half the equivalent GPT-5.6 pricing, with Sol at $2 input and $10 output per million tokens and Luna at $0.10 and $0.50, plus a roughly 90 percent discount on cached input reads.

Does a cheaper price mean lower quality?

Not necessarily, and OpenAI does not claim that. The company says Sol and Luna use training methods similar to Astra’s, which would mean the price drop reflects efficiency and tier positioning rather than a quality cut. That claim has not been independently tested on ordinary writing tasks as of this release.

Should I switch my team’s default model because of this release?

Only after testing it on your own writing rather than the vendor’s benchmarks. See how to benchmark a local model on your own writing for an approach that works the same way against a hosted model.

Download Wrivio for Windows to keep a stable set of rewrite Contexts that work the same way no matter which model or tier is behind them this month.