1300 166 138 Tell us what you need
News / OpenAI

OpenAI halves GPT 6 Sol prices on Opus 5.5 launch day

By Monika Tantau • Published 2 months ago
ChatGPT app icon in a computer dock with the cursor over it, for OpenAI's GPT 6 Sol and GPT 6 Luna launch

Key takeaways

• OpenAI released GPT 6 Sol and GPT 6 Luna on 22 September, hours after Claude Opus 5.5
• Sol costs $2.81 per million input tokens and $14.04 output, exactly half the GPT 5.6 Sol rate
• Luna costs $0.14 and $0.70, a hundred times less per token than GPT 6 Astra
• There is no GPT 6 Terra. The family is Astra, Sol and Luna, and Sol now covers the middle

Most Read

OpenAI released two models on 22 September, GPT 6 Sol and GPT 6 Luna, and priced both at half or less of what the models they replace cost. Sol runs at $2.81 per million input tokens and $14.04 per million output. Luna, the smaller of the two, costs $0.14 and $0.70 for the same volumes.

The timing is as much of the story as the arithmetic. Anthropic had taken the top of the independent leaderboard that morning with Claude Opus 5.5 and trimmed its own rates by 20%. OpenAI answered by halving the price of the models most teams actually run in production, rather than the one at the top of its range.

What OpenAI actually shipped

Sol is the working model of the pair. OpenAI describes it as built to power complex coding and agentic workflows, and points at repeat work such as building features, reviewing code, debugging and analysing data. Luna is aimed the other way, at focused high volume jobs like summarising documents, pulling fields out of them and answering short questions.

Both carry the same envelope as the rest of the family: a context window of 1,050,000 tokens and a ceiling of 128,000 tokens of output. Sol knows the world to 20 April 2026 and Luna to 18 May 2026, according to OpenAI documentation. Both accept text and images and return text.

The price cut is the part that is not incremental. GPT 6 Sol is exactly half of GPT 5.6 Sol on both sides of the meter. Luna is half on input and roughly 58% below on output. Cached input on either model is billed at a tenth of the uncached rate, which matters more than it sounds for anything that resends the same instructions all day.

Luna is a hundred times cheaper than Astra

The spread inside a single model family is now wider than the gap between most vendors. GPT 6 Astra, released earlier this month and still the most capable of the three, costs a hundred times more per token than Luna on both input and output. That is a real choice rather than a marketing tier, and it only pays off if the work is sorted before it is sent.

Output price per million tokens, Australian dollars

GPT 6 Astra $70.20
GPT 5.6 Sol $28.08
GPT 5.6 Terra $16.85
GPT 6 Sol $14.04
GPT 5.6 Luna $1.69
GPT 6 Luna $0.70

Darker bars are the GPT 6 family. List prices from OpenAI model documentation, converted at the Reserve Bank rate for 22 September 2026.

Read down that chart and the shape of the release is clear. GPT 6 Sol has landed below the model that used to be the mid tier of the previous family, while doing the work the flagship of that family was bought for. Our breakdown of the GPT 5.6 lineup now describes a set of prices that no longer exists at the top.

The long prompt surcharge did not go away

One thing did not get cheaper. Both models keep the rule that arrived with Astra: a request carrying more than 272,000 input tokens is billed at double the input and cache rates and one and a half times the output rate, and the multiplier applies to the whole request rather than the excess.

That matters because the context window is 1,050,000 tokens. The headline number and the number where the meter changes are nearly four times apart, so a pipeline that quietly grows its prompt can cross the line without anything in the response saying so. Batch and Flex still run at half rate, and fast mode at double.

Where you can use them today

Both models are live in ChatGPT Work and in Codex for Plus, Pro, Business, Enterprise and Edu subscribers, and Luna reaches Free and Go users through the desktop app. In GitHub Copilot, Sol arrived on Pro+, Max, Business and Enterprise plans and Luna on those plus Pro, billed as usage.

The rollout is gradual, so a team that cannot see the models yet is not misconfigured. Independent scores arrived within hours: on the Artificial Analysis index that put Claude Opus 5.5 in first place the same day, GPT 6 Sol at maximum effort scores 47.53, half a point above GPT 5.6 Sol, and neither model has a measured cost per task yet.

Why Australian businesses should care

The models that decide whether an automation project is affordable are almost never the flagship. They are the ones underneath, and both of those just halved. A summarising job that processes 5,000 documents a month at roughly 2,000 output tokens each runs to about $7 a month in output charges on Luna, against $17 on the model it replaces.

That is the arithmetic that moves a pilot into production. Work that involves reading a lot of text and writing a little, which is most document automation, is now cheap enough that the model is no longer the line item anyone argues about. What remains expensive is the part around it: the extraction rules, the exceptions, and the checking.

OpenAI bills in United States dollars, so what a local team pays each month moves with the exchange rate rather than sitting still. Every figure here is converted at the Reserve Bank rate published on 22 September 2026, when one Australian dollar bought 0.7123 United States dollars.

For anyone already weighing the two ends of the market, the comparison has moved again. Our piece on Astra against Claude Fable 5.1 was written when the cheapest way to run OpenAI’s newest generation was ten times what it is this week.

Common questions

What is the difference between Astra, Sol and Luna?

Astra is the most capable and the most expensive at $14.04 and $70.20 per million tokens. Sol is built for coding and agent work at $2.81 and $14.04. Luna is the high volume option at $0.14 and $0.70. All three share the same context window and output ceiling.

Which one should we use for code?

Sol. OpenAI positions it for complex coding and agentic workflows, and reports that it makes about half as many factual mistakes as its predecessor. Astra remains the one to reach for on the hardest problems, at five times the input price.

What happened to GPT 6 Terra?

There is no GPT 6 Terra. The 5.6 family ran Sol, Terra and Luna; the 6 family runs Astra, Sol and Luna. In practice Sol absorbed the middle: it matches GPT 5.6 Terra’s input price of $2.81 and comes in below it on output.

What do they cost in Australian dollars?

Sol is $2.81 per million input tokens and $14.04 per million output. Luna is $0.14 and $0.70. Cached input on either is a tenth of the uncached rate. OpenAI bills in United States dollars, so those figures move with the exchange rate.

A floor that keeps dropping

Two of the largest laboratories in the industry cut prices on the same day, neither of them under obvious commercial pressure to do so. Anthropic attached its cut to a new leader and OpenAI attached its own to the tier below the leader, which are different arguments about where the competition actually is.

What both admit is that serving these models is getting cheaper faster than the public rates had been showing. For a business here that has already built something and is watching what it costs to run, that is a better week than any benchmark result has delivered this year.

The teams that gain most are the ones with a workload already instrumented, because they can change the model and read the difference in a fortnight. For everyone still deciding what to automate, the price of the model was never what was holding it up.

Scroll to Top