GPT-6 Sol and Luna landed on September 22, 2026, and the interesting number is not on any benchmark chart. It is on the pricing page.
GPT-6 Sol costs $2.00 per million input tokens and $10.00 output. The model it replaces, GPT-5.6 Sol, is $4.00 and $20.00. That is half, on both sides of the meter.
The capability story is thinner. The only independent measurement published on launch day puts GPT-6 Sol one point above its predecessor, and GPT-6 Luna two points below its own.
We covered what GPT-6 Astra actually shipped on September 11 and the gap between its ARC-AGI-3 claim and a neutral harness three days later. This is the same exercise on a launch that is twelve hours old.
What shipped, and where you can actually use it
Two models, available immediately in the API as gpt-6-sol and gpt-6-luna. Neither carries a dated snapshot suffix; the model ID and the snapshot ID are the same string.
Availability is narrower than the headlines imply. From OpenAI’s announcement, verbatim: "GPT-6 Sol and GPT-6 Luna are available in ChatGPT Work and Codex starting today for all Plus, Pro, Business, Enterprise, and Edu users. Free and Go users can access GPT-6 Luna in the desktop app. These models are not yet available in Chat."
That last sentence is the one to carry into a client conversation. Staff on ChatGPT Plus who open the normal chat interface looking for the new models will not find them today. This is ChatGPT Work and Codex on day one. OpenAI also says the rollout is gradual across the day, so absence in a UI right now proves nothing either way.
Both models are already in GitHub Copilot, per its changelog, with Sol on Pro+ and above and Luna from Pro upward, enabled by default for Business and Enterprise unless an admin turns them off.
Both carry a 1,050,000-token context window and 128,000 maximum output tokens, matching Astra. Both support reasoning efforts from none through max. Astra, notably, does not support none, which matters if you are moving work down from it.
The prices are the actual news
Set out in full, because this is where the decision gets made.
| Model | Input | Cached input | Output |
|---|---|---|---|
| GPT-6 Astra | $10.00 | $1.00 | $50.00 |
| GPT-6 Sol | $2.00 | $0.20 | $10.00 |
| GPT-6 Luna | $0.10 | $0.01 | $0.50 |
| GPT-5.6 Sol | $4.00 | $0.40 | $20.00 |
| GPT-5.6 Terra | $2.00 | $0.20 | $12.00 |
| GPT-5.6 Luna | $0.20 | $0.02 | $1.20 |
GPT-6 Sol is one fifth of Astra’s price on both input and output. Luna is one hundredth.
OpenAI describes this as "reducing API prices for Sol and Luna by 50% compared with their GPT-5.6 promotional pricing." Sol is exactly 50% on both. Luna’s input is 50% but its output falls from $1.20 to $0.50, which is 58%, better than advertised. Worth noting the comparison base is described as promotional pricing, and GPT-5.6 had at least two prior price actions this year, so the 50% is measured against a baseline that had already moved.
Three cost lines that do not appear in the headline and will appear on an invoice:
The long-context cliff. Requests above 272,000 input tokens are billed at 2x input and 1.5x output for the entire request, not just the overflow. A 280,000-token call on Sol costs $4.00 and $15.00 per million across every token in it. Chunking below that line is a straight halving.
Cache writes are not free. $2.50 per million on Sol, $0.125 on Luna. Prompt caching saves a great deal on reads and costs something on writes, and cost models routinely omit the second half.
Batch and Flex are 50% of standard. On Luna that is $0.05 input and $0.25 output. For overnight bulk work, archive tagging, metadata extraction, back-catalog summaries, that is close to free and it is the single biggest lever available to a budget-constrained client.
Regional processing adds 10%. EU data residency is available only with standard processing, so you cannot combine the two.
The independent numbers say capability is flat
Here is the part that has not made it into most of today’s coverage.
Artificial Analysis ran its own evaluation suite on launch day. It is not a reproduction of OpenAI’s benchmarks, it is a different set of tests, and as of this writing it is the only independent measurement of these models that exists.
On its Intelligence Index at max effort: GPT-6 Sol scores 48. GPT-5.6 Sol scores 47. For reference in the same index, GPT-6 Astra is 53, Claude Fable 5.1 is 53, and Claude Opus 5.5 is 58.
On cost per task, GPT-6 Sol runs $1.06 against GPT-5.6 Sol’s $1.99, about 47% less.
Artificial Analysis’s own summary is that scores "remain level with GPT-5.6, with progress in some evaluations and regressions in others."
One point of intelligence for half the money is a good trade and it is not a new generation of capability. Read against the price table, the honest description of this launch is that OpenAI made its mid and low tiers substantially cheaper and left the frontier where Astra put it three weeks ago.
Luna went backwards on two measures
The Luna result deserves separating out, because it is the one that could cost somebody something.
On Artificial Analysis’s Intelligence Index, GPT-6 Luna scores 37, the same as GPT-5.6 Luna. On its Coding Agent Index, GPT-6 Luna scores 41 against GPT-5.6 Luna’s 43, a two-point regression. Luna also reportedly lost around 45 Elo on AA-Briefcase v1.1, and both new models scored down on GDPval-AA v2.1, which measures real-world economic work.
Luna is the tier that does volume work. Extraction, classification, summarization, tagging, the jobs you run ten thousand times rather than ten. That is precisely the workload where a two-point drop compounds and where nobody is watching closely enough to notice.
The price cut is real and the capability is not obviously better. For an existing Luna workload, that is a reason to run your own evaluation before switching, not a reason to switch on the strength of the price.
Every OpenAI figure is self-reported, against a baseline that aged in 90 minutes
OpenAI published five benchmark comparisons. All are self-reported. None has been independently reproduced today, and ARC Prize’s leaderboard carries no entry for either model.
- AutomationBench 1.0.6: Sol 33.2% at xhigh, against Claude Opus 5 at max, 26.9%.
- Agents’ Last Exam V1: Sol 56.4% at max, described as above Claude Opus 5’s highest score at 60% lower cost per task.
- FrontierCode 1.1 Main: Sol "matches Claude Fable 5.1 xhigh at much lower cost." No number given.
- DeepSWE v1.1: Sol 68.8% at max against Claude Fable 5 xhigh at 69.9%. Sol is behind here; OpenAI frames it as within 1.1 points.
- OSWorld 2.0 offline: Sol 60.5% at xhigh against Claude Opus 5 at medium, 60.3%.
Two cautions. The Agents’ Last Exam sentence is easy to misread: the 60% is the cost reduction, not Opus 5’s score, and Opus 5’s actual score on that evaluation is not recoverable from the announcement. And the DeepSWE row is a loss presented as a near-tie.
Then the thing that undercuts the whole comparison set. TechCrunch reports that Anthropic shipped Claude Opus 5.5 roughly ninety minutes before OpenAI’s announcement. Every OpenAI comparison above is against Opus 5 or Fable 5.1. Artificial Analysis puts Opus 5.5 at 58 on its index against GPT-6 Sol’s 48.
That is not an accusation of bad faith. Benchmark tables are built weeks ahead and a competitor’s release schedule is not OpenAI’s to control. It does mean the competitive claims in this announcement describe a landscape that stopped existing an hour and a half before they were published, and anyone quoting them into a client deck should know that. We set out how to read a launch-day benchmark table in vendor benchmark scores are not leaderboard results.
The factuality claim is narrower than the headline
Coverage has run with "fewer mistakes." OpenAI’s actual sentence:
"On our internal factuality evaluation, which is based on de-identified real-world conversations where users flagged mistakes by our models, GPT-6 Sol makes about half as many mistakes as its predecessor, approaching Astra-level reliability."
Unpacking that: it is an internal, proprietary evaluation that nobody outside OpenAI can run. It is built from user-flagged mistakes in real conversations rather than a standard hallucination benchmark. The comparison is against GPT-5.6 Sol, not against a competitor and not against Astra. And it is about Sol only. We found no equivalent claim for Luna.
There is no absolute error rate, no figure on any named public benchmark such as SimpleQA or PersonQA, and no published methodology. Halving your own unpublished internal metric is a real engineering result and it is not a number anyone else can check.
Where Astra sits now, and the hole where Terra was
The lineup question is the one clients will ask, so here is OpenAI’s own framing, quoted: "GPT-6 Astra continues to be our best model across the board. Choose it when you want the best results and an uncompromising experience."
So it is a straight ladder. Astra above Sol above Luna, on capability and on price, not three points on different axes. Astra is not deprecated and not demoted. What changed is that it is no longer the only way to buy GPT-6, and its five-times premium over Sol now has to justify itself against a five-point gap on the one independent index available.
There is no GPT-6 Terra. The models index lists exactly three GPT-6 models. Terra existed in GPT-5.6 as the mid-tier at $2 input and $12 output. GPT-6 Sol now occupies that same $2 input price while being the upper of the two new models. A community thread asking about GPT-6 Terra has no reply from OpenAI. We are not going to tell you it is coming or that it is not, because OpenAI has said nothing either way.
One oddity worth recording rather than explaining: the published knowledge cutoff for Sol is April 20, 2026 and for Luna is May 18, 2026. The cheaper model has the later cutoff. That is what the model pages say and we found no explanation for it.
Safety documentation is not a separate publication. It is an appendix added to the GPT-6 Astra system card on September 22. If a nonprofit client has a procurement or ethics review, that appendix is the document to hand over.
Rate limits, endpoints and the things that block a build
Neither model is gated behind a spend tier. Both are available from API Tier 1, which matters if you are standing up a small client integration rather than moving an existing one.
The ceilings differ between them in a way that is not obvious from the price table. GPT-6 Sol runs 500 requests and 500,000 tokens per minute at Tier 1, rising to 15,000 requests and 40 million tokens at Tier 5. GPT-6 Luna starts identically at Tier 1 but climbs further and faster, reaching 30,000 requests and 180 million tokens per minute at Tier 5. Luna’s batch queue and throughput headroom are built for the volume work it is priced for.
Sol supports Chat Completions, Responses and Batch. It does not support Realtime, Assistants, fine-tuning, embeddings, audio or video, or the legacy Completions endpoint. Tool support is broad: web search, file search, image generation, code interpreter, hosted shell, apply patch, skills, computer use and MCP, plus streaming, structured outputs, function calling, image input and prompt caching.
One honest gap. We could not get a clean read on Luna’s endpoint support. The documentation we extracted listed Realtime, fine-tuning, Assistants and embeddings as supported, which is inconsistent with Sol and implausible for a text model in this family. Check Luna’s model page directly before you design around any of those. Its modality support we are confident in: text and image in, text out, no audio or video.
For background on the tier names themselves and where they came from, see our explainer on the GPT-5.6 family.
What to do
If you are running GPT-5.6 Sol, move to GPT-6 Sol and re-test. Half the cost, capability level or a point better on the only independent measure. This is the closest thing to free money in the announcement.
If you are running GPT-5.6 Luna, do not move on price alone. Run your own evaluation first. The independent numbers show a coding regression and a drop on real-world work, and Luna’s workloads are high volume enough that a small regression is expensive and quiet.
Check what your clients see. These are in ChatGPT Work and Codex, not consumer Chat. A client whose staff have Plus seats will not find them in the usual interface today.
Chunk below 272,000 tokens and budget for cache writes. Both are easy to miss and both show up on the bill.
Run bulk work through Batch or Flex. Half price again, and on Luna that puts high-volume back-catalog work into the range where cost stops being the constraint.
One more note on timing. OpenAI DevDay is September 29, seven days out. Shipping these models a week ahead of it rather than on the stage suggests the DevDay agenda is something other than these two models. That is inference rather than reporting, and we will find out on the day.
Frequently Asked Questions
What are GPT-6 Sol and Luna?
Two models OpenAI released on September 22, 2026, available in the API as `gpt-6-sol` and `gpt-6-luna`. They sit below GPT-6 Astra on a capability and price ladder and cost roughly half what the GPT-5.6 models they succeed cost.
What do they cost?
GPT-6 Sol is $2.00 per million input tokens and $10.00 output. GPT-6 Luna is $0.10 input and $0.50 output. Cached input is $0.20 and $0.01. For comparison, GPT-6 Astra is $10.00 and $50.00.
Are they actually better than GPT-5.6?
Barely, on the one independent measurement available. Artificial Analysis scores GPT-6 Sol at 48 on its Intelligence Index against GPT-5.6 Sol’s 47, and GPT-6 Luna level with GPT-5.6 Luna at 37 while dropping two points on its Coding Agent Index. The launch is better described as a price cut than a capability jump.
Has anyone verified OpenAI’s benchmark numbers?
No. All five published comparisons are self-reported and none has been independently reproduced. Artificial Analysis ran a different suite rather than a reproduction, and ARC Prize’s leaderboard has no entry for either model.
Why does the Claude comparison matter?
OpenAI benchmarked against Claude Opus 5 and Fable 5.1. Anthropic released Claude Opus 5.5 roughly ninety minutes before OpenAI’s announcement, and Artificial Analysis scores Opus 5.5 at 58 against GPT-6 Sol’s 48. The comparison set describes a landscape that had already changed.
What is the “fewer mistakes” claim?
OpenAI says GPT-6 Sol makes about half as many mistakes as its predecessor on an internal factuality evaluation built from de-identified conversations where users flagged errors. It is proprietary, unpublished, applies to Sol only, and is measured against GPT-5.6 Sol rather than any competitor.
Can I use them in ChatGPT?
In ChatGPT Work and Codex, yes, for Plus, Pro, Business, Enterprise and Edu users. Free and Go users get Luna in the desktop app. OpenAI states they are not yet available in Chat.
What is the context window?
1,050,000 tokens with 128,000 maximum output for both, matching GPT-6 Astra. Requests above 272,000 input tokens are billed at 2x input and 1.5x output across the entire request.
Is there a GPT-6 Terra?
No. The GPT-6 family is Astra, Sol and Luna. Terra existed in GPT-5.6 as the mid-tier. OpenAI has not said whether a GPT-6 Terra is planned.
Is GPT-6 Astra being retired?
No. OpenAI describes it as “our best model across the board” and it is not on the deprecations page. What changed is that it is no longer the only GPT-6 option.
Are the GPT-5.6 models deprecated now?
No. They do not appear as deprecated anywhere on OpenAI’s deprecations page, and neither the announcement nor the model guidance says GPT-5.6 is retiring.
Does this change the October 23 shutdown?
Not as of today. The deprecations page has no entry dated September 22 and still recommends migrating retired models to GPT-5.6 Sol, Terra and Luna, which are now more expensive than GPT-6 Sol at every tier.
