On September 22, 2026, OpenAI announced GPT-6 Sol and GPT-6 Luna (Introducing GPT-6 Sol and Luna).
GPT-6 Astra had arrived earlier in the same month, so the pace of model releases is still very fast.
What interests me most this time is not simply that another pair of models has appeared, but that OpenAI has raised capability while pushing down the cost of using that capability.
In my earlier article, GPT-6 Astra Is Here. For Daily Coding, GPT-5.6 Luna May Still Be Enough, I argued that Astra made sense for difficult work while GPT-5.6 Luna remained attractive for everyday development because of its cost efficiency.
GPT-6 Sol and Luna push that idea further.
GPT-6 Sol and Luna are cheaper than the 5.6 generation
OpenAI describes GPT-6 Sol and Luna as bringing progress from Astra into faster, lower-cost models.
The clearest change is API pricing.
| Model | Input / 1M tokens | Output / 1M tokens |
|---|---|---|
| GPT-6 Astra | $10.00 | $50.00 |
| GPT-6 Sol | $2.00 | $10.00 |
| GPT-6 Luna | $0.10 | $0.50 |
| GPT-5.6 Sol (promotional price) | $4.00 | $20.00 |
| GPT-5.6 Terra | $2.00 | $12.00 |
| GPT-5.6 Luna (promotional price) | $0.20 | $1.20 |
GPT-6 Sol cuts both input and output pricing by 50% compared with the promotional GPT-5.6 Sol rates.
Luna halves the input price and drops output pricing from $1.20 to $0.50 per million tokens.
OpenAI attributes these reductions to improvements in inference processing and cache efficiency. Current rates are also listed on the OpenAI API Pricing page.
This is not just a cheaper tier arriving beneath the flagship. The models got stronger while their token prices fell.
Sol is not just a cheaper 5.6 Sol
The lower price does not appear to come from simply scaling capability back.
In OpenAI's internal factuality evaluation, GPT-6 Sol is described as producing roughly half as many errors as GPT-5.6 Sol.
For coding, OpenAI also reports a substantial improvement on FrontierCode, an evaluation of whether real codebase changes are good enough to merge.
GPT-6 Sol with max reasoning also reaches 68.8% on DeepSWE v1.1.
Benchmarks do not map perfectly to day-to-day use, but this does not look like a case where the price was cut by making the model weaker. It feels closer to bringing capability above GPT-5.6 Sol down toward GPT-5.6 Terra pricing.
Luna looks even more useful for agent workloads
GPT-6 Luna may be the more interesting model for high-volume agent work.
On OpenAI's AutomationBench results, higher reasoning improves performance by 5.4 points over GPT-5.6 Luna while reducing per-task cost by 58%.
On OSWorld 2.0, OpenAI says GPT-6 Luna with max reasoning outperforms GPT-5.6 Sol with medium reasoning at roughly one-tenth the cost.
GPT-6 Luna also supports none, low, medium, high, xhigh, and max reasoning effort through the API (GPT-6 Luna Model).
That is a useful shape for agent systems. Code search, classification, small edits, test additions, and bounded investigations do not all need the most capable model available, so it is natural to run many cheap tasks on Luna and move only the difficult parts upward.
My current split would look something like this:
- GPT-6 Luna: exploration, classification, small edits, high-volume subtasks
- GPT-6 Sol: normal implementation, review, and moderately complex agent work
- GPT-6 Astra: architecture, difficult investigations, unclear failures, and long end-to-end work
I previously thought "Luna for daily work, Astra for the hard parts" was already a reasonable setup. At this price, GPT-6 Sol now looks like a much stronger candidate for the everyday default.
What 100k input and 20k output looks like
Assume one uncached task consumes 100,000 input tokens and 20,000 output tokens.
| Model | Estimated token cost |
|---|---|
| GPT-6 Astra | ~$2.00 |
| GPT-6 Sol | ~$0.40 |
| GPT-6 Luna | ~$0.02 |
| GPT-5.6 Sol (promotional price) | ~$0.80 |
| GPT-5.6 Terra | ~$0.44 |
| GPT-5.6 Luna (promotional price) | ~$0.044 |
GPT-6 Sol is about one-fifth the cost of Astra and almost the same price as GPT-5.6 Terra.
Luna is roughly two cents under the same simplified assumption.
Actual bills vary with reasoning tokens, tool calls, caching, long-context pricing, and output size, but the gap is still large enough to matter when an agent may retry several times.
Once AI is capable enough, the next practical question stops being only "can I hand this task to a model?" and becomes "can I afford to hand this task to a model every time?" GPT-6 moves that boundary noticeably.
Where did Terra go?
The naming has become more confusing at the same time.
GPT-5.6 had Sol, Terra, and Luna. OpenAI described them as "durable capability tiers," which sounded like model tiers intended to persist independently of the generation number (GPT-5.6).
With GPT-6, the current main lineup is Astra, Sol, and Luna.
Terra is not in the current GPT-6 model list, although OpenAI has not officially said that Terra was discontinued or folded into Sol, so I would not treat that as a confirmed transition.
From a user's point of view, however, the roles can look something like this:
- GPT-5.6 Sol → closer to the flagship position now occupied by GPT-6 Astra
- GPT-5.6 Terra → closer to the everyday position now occupied by GPT-6 Sol
- GPT-5.6 Luna → GPT-6 Luna
GPT-6 Sol is more capable than GPT-5.6 Sol, so this is not a literal successor mapping.
There is another small naming detail that bothers me: GPT-5.6 is followed by GPT-6, not GPT-6.0. If I read the names like software versions, I instinctively want to write "GPT-6.0," but OpenAI consistently uses "GPT-6," so that is the official form used here as well.
That little discomfort is part of the same problem: the names no longer make generation and capability tier especially easy to infer.
Using the same "Sol" label while its apparent place in the lineup shifts between generations is especially awkward. In Claude terms, it would feel a bit like a previous-generation Opus role showing up under Sonnet in the next generation.
As codenames, though, Sol / Terra / Luna / Astra are actually pretty memorable. You can loosely imagine Astra > Sol > Terra > Luna, with the larger real-world object corresponding to the more capable model.
That is a nice mnemonic. I am less convinced it works as the customer-facing product hierarchy.
For external product names, something like Fast / Standard / Pro or S / M / L would be easier to understand at a glance.
Model naming is not only branding; it is also UX. I would like to see this cleaned up.
For Plus users, GPT-6 Sol is the model to watch
I use ChatGPT Plus and spend a lot of time talking with GPT-5.6 Sol in normal Chat, so GPT-6 Sol interests me more than Astra for everyday use.
At article-writing time, GPT-6 Sol and Luna are available in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu, while OpenAI says they are not yet available in normal Chat.
That means Plus users cannot simply move their normal GPT-5.6 Sol conversations to GPT-6 Sol yet.
GPT-6 Astra, by contrast, was announced for gradual rollout to ChatGPT users including Plus.
Given the balance of price and capability, GPT-6 Sol looks like a very natural candidate for the main model in normal Chat, but that is my inference; OpenAI has not announced when GPT-6 Sol will arrive there.
That is the next part I want to watch as a Plus user.
Codex Default may be a step toward hiding model choice
Around the same time as this release, Codex added a "Default" option to model selection.
OpenAI's help page describes Default as a way to combine model and reasoning level into a single slider (ChatGPT Work and Codex). If you manually choose a specific model, Codex remembers that choice.
The official description available at article-writing time does not make it clear whether Default dynamically routes each task between Luna, Sol, and Astra, or whether it simply maps the slider to predetermined model-and-reasoning combinations.
What it does suggest is a move away from forcing users to think about model names every time they want to trade cost for capability.
OpenAI's own Model selection guide follows the same general logic: small edits and bounded problems can start on Luna Low, while implementation from clear requirements can use Luna Medium.
If Default eventually absorbs more of that decision-making, the increasingly confusing model names matter less at the UI layer.
The more SKUs there are, the less appealing it is to make users choose one for every task. I would rather select something closer to speed, cost, and thinking depth, then let the system move toward an appropriate model behind the scenes.
The current Default option is not officially documented as that kind of dynamic router yet, so this is another area worth watching.
Falling cost matters more than a faster release cycle
GPT-5.6 arrived in July, GPT-6 Astra in early September, and Sol / Luna later in September.
There is still a broader debate about how quickly frontier AI development should move, while the product release cadence remains very fast. Following new models every few months is already a lot of work, and the naming does not make it easier.
I still like the direction of this release because the interesting part is not only a few more benchmark points. OpenAI is also improving cache and inference efficiency so the same budget buys more useful model work.
That matters even more as agents run longer and spawn more subtasks.
Research on flagship models such as Astra still matters, but feeding those gains back into Sol and Luna so they become cheaper and widely usable is the healthier half of the cycle.
At least for me, "I can use this level of capability repeatedly at this price" is a more exciting improvement than "the model got a little smarter again."
Comments