GPT-6 Sol and Luna: prices halved, the same evening as Opus 5.5

The essentials in 30 seconds ⚡
OpenAI released GPT-6 Sol and GPT-6 Luna on the evening of 22 September, two models placed under the flagship Astra. Sol is priced at 2 dollars per million input tokens and 10 for output; Luna at 0.10 and 0.50. That is roughly half the rates of their predecessors, and the company confirms these are permanent prices. The release comes a few hours after Claude Opus 5.5, also announced with a sharp price cut.

Two competing labs led their announcement with price, on the same evening. It is the most significant fact of the week, and it says more than the scores do.

The rates, side by side

Model	Input / M	Output / M	Change
GPT-6 Sol	$2.00	$10.00	−50% / −50%
GPT-5.6 Sol (previous)	$4.00	$20.00	promotional pricing
GPT-6 Luna	$0.10	$0.50	−50% / −58%
GPT-5.6 Luna (previous)	$0.20	$1.20	—
Claude Opus 5.5 (same evening)	$4.00	$20.00	−20% advertised, −40% actual

Two details are worth noting.

Sol's prices match those of Claude Sonnet 5 exactly on every line published by the two providers: $2 for input, $0.20 for a cache read, $10 for output. That alignment is probably not a coincidence.

The comparison is made against a temporary rate. The 50% drop is measured against the promotional rate for GPT-5.6 Sol, guaranteed only until 21 November 2026. Nobody knows what the older model will cost afterwards. This is exactly the trap we described with the end of Sonnet 5's promotional rate: a reduction announced against an introductory price is not a reduction against a market price.

The asterisk the press release doesn't highlight

Here is the most interesting point, and it requires reading the data behind the charts rather than the points chosen for the announcement.

On two of the six evaluations highlighted — an agentic coding test and a computer-use test —, GPT-6 Sol's best score is lower than GPT-5.6 Sol's: around 68.8% versus 72.7%, and 64.4% versus 66.2%. On those same two tests, Claude Opus 5 also did better.

In other words: on certain tasks, the newer model performs less well than the one it replaces. This is not an anomaly, and OpenAI does not hide it — the company explicitly positions these models on efficiency rather than on a new capability ceiling. But it is information that "prices halved" does not let you guess.

What this really means 📉
A cheaper model that scores slightly lower is not a bad model: it is a trade-off, and it can be an excellent one. On a business workflow automation test, Sol reaches 33.2% at maximum effort for 27 cents per task — ahead of Astra at low effort, which gets 30.3% while costing nearly four times more, and ahead of Opus 5 at maximum effort, which gets 26.9% while costing eleven times more. The right criterion is therefore not the score but the score-cost pair at a given effort level. A model that loses four points and divides its bill by eleven wins in most real-world uses.

Luna, the model nobody will talk about

It is probably the most consequential product of the announcement, and it is the least spectacular.

At 0.10 dollars per million input tokens, Luna costs about a tenth of Claude Haiku 4.5, and a fraction of what xAI's latest model, released the day before, charges.

At that level, an entire line of spending disappears. Classification, field extraction, intent detection, filtering: everything we filed under the "under 3 billion parameters" category in our size guide becomes billable at negligible amounts.

It is the same shift we described for transcription at ten cents an hour: when an operation becomes cheap enough, it is not its volume that increases, it is its nature that changes. You stop making trade-offs and process everything.

The specifications

The two models share a context window of around 1.05 million tokens and up to 128,000 output tokens, with reasoning settings ranging from none to maximum. They accept text and images as input, produce text, and come with the usual agent tools.

The knowledge cut-off dates are 20 April 2026 for Sol and 18 May for Luna. The identifiers are gpt-6-sol and gpt-6-luna.

On deployment, both arrive in the workspace and the coding tool for paid plans; Luna is available to free accounts via the desktop app. Neither is offered in the ordinary chat interface at launch.

OpenAI also claims roughly half as many factual errors for Sol as for its predecessor, on an internal evaluation.

What the simultaneity tells us

Two competitors leading their announcement with price, on the same evening, a few hours apart: that is the structural fact of the week.

It confirms what we have been describing for a month. Competition is no longer about the capability ceiling — Astra remains at the top for one, Fable 5.1 for the other, and neither has moved this week. It is about cost per task completed, because that is what decides whether an agent-based product is profitable.

We wrote as much about the startup valued at 1.4 billion that trains no model: the base layer is becoming commoditised and value is moving up to what is built on top of it. A price war between the two main providers is the financial expression of that commoditisation.

What to take away

For a user, the week is excellent: two major providers cut their prices on the same day, and the competition on inference that we documented with the Qualcomm-Amazon deal is starting to have effects all the way down to the bill.

When choosing, the method does not change and is becoming more worthwhile than ever. Take your reference cases, test at the same level of effort, and compare cost per successful task — not the score, not the advertised rate.

And bear in mind the asterisk: a newer model is not necessarily better on your task. On two tests, this one does worse than the one it replaces. That is the subject we cover right after in a dedicated article.