GPT-6 Sol and Luna: Half the Price, and Where Each One Fits

GPT-6 Sol and Luna are OpenAI's two cheaper GPT-6 models, released on September 22, 2026, at half the price of the GPT-5.6 models they replace. Independent tests show the savings are real: Sol now costs $1.06 per task at its top setting, down from $1.99. But capability barely moved. Sol scores 48 against 47 before, and Luna matches its predecessor at 37.

Key Takeaways

• OpenAI released GPT-6 Sol and GPT-6 Luna on September 22, 2026, at half the list price of GPT-5.6 Sol and Luna.
• GPT-6 Sol costs $2 input and $10 output per million tokens. GPT-6 Luna costs $0.10 and $0.50.
• On independent tests, Sol costs about 47% less per task than GPT-5.6 Sol and runs about 55% faster, but scores just 1 point higher.
• Luna scores the same as GPT-5.6 Luna for less than half the cost per task, at about 154 tokens per second.
• Both have weak spots. Sol slipped on one knowledge-work test, and Luna scores near zero on a factual-recall test that penalizes wrong answers.

What Did OpenAI Release With GPT-6 Sol and Luna?

Two cheaper, faster members of the GPT-6 family. OpenAI says they "build on the advances behind GPT-6 Astra," its flagship, which we covered in our GPT-6 Astra analysis.

Reported fact: OpenAI announced the two models on September 22, 2026, according to TechCrunch. The launch came about 90 minutes after Anthropic released Claude Opus 5.5, which we covered in our Claude Opus 5.5 analysis. OpenAI positions Sol for complex tasks, including coding, and Luna for high-volume work such as summarizing documents, extracting information and answering questions.

Both are available in the API as gpt-6-sol and gpt-6-luna. They're also in ChatGPT for paid plans and in Codex, and Luna is available to Free and Go users in the desktop app. GitHub says both are in Copilot, with Sol on Pro+, Max, Business and Enterprise, and Luna also on Pro.

Here's the pricing, per million tokens, as reported by VentureBeat:

Price per million tokens	GPT-6 Sol	GPT-5.6 Sol	GPT-6 Luna	GPT-5.6 Luna
Input	$2.00	$4.00	$0.10	$0.20
Output	$10.00	$20.00	$0.50	$1.20
Cached input	$0.20	—	$0.01	—

OpenAI's spokesperson told reporters the lower prices are permanent, not a promotion. Simon Willison called Luna "one of the cheapest models OpenAI have ever released."

One detail to check: Artificial Analysis lists GPT-6 Sol's context window at about 872,000 tokens, down from 1 million for GPT-5.6 Sol. Luna keeps 1 million.

How Much Cheaper Are GPT-6 Sol and Luna per Task?

A lot. For Sol, the per-task savings are close to the price cut, because it uses about the same number of output tokens as before, around 31,000 per task against 29,000.

Here's how the new and old models compare at their top settings on the Artificial Analysis Intelligence Index, version 4.3.2:

Model (top setting)	Intelligence Index	Cost per task	Output speed
GPT-6 Sol	48	$1.06	About 125 tokens/s
GPT-5.6 Sol	47	$1.99	About 80 tokens/s
GPT-6 Luna	37	$0.07	About 154 tokens/s
GPT-5.6 Luna	37	$0.18	About 142 tokens/s

So GPT-6 Sol costs about 47% less per task and runs about 55% faster. GPT-6 Luna costs about 61% less per task for the same score.

Lower settings cut costs much further. Artificial Analysis scores GPT-6 Sol at 43 on high for $0.37 per task, and 40 on medium for $0.25. Luna scores 32 on high for $0.03 per task. For many jobs, those settings are the real sweet spot.

What Does the Price Cut Mean for a Monthly Bill?

Half, before any other savings. Here's a simple example with round numbers. Say an app sends 50 million input tokens and gets 10 million output tokens back each month, at list prices:

• GPT-5.6 Sol: $200 for input plus $200 for output, or $400 a month.
• GPT-6 Sol: $100 for input plus $100 for output, or $200 a month.
• GPT-6 Luna: $5 for input plus $5 for output, or $10 a month.

Caching changes the math further. If 80% of those input tokens are cached prompts, GPT-6 Sol's input bill falls from $100 to $28, because cached input costs $0.20 per million instead of $2. The total drops to about $128 a month. At higher volumes, those differences grow quickly, which is why caching is worth setting up early.

Two cautions apply. These figures assume the same token counts, which holds for Sol but should be checked on your own work. And a cheaper model that fails more tasks can cost more in retries and human review than it saves on tokens.

Is GPT-6 Sol Smarter Than GPT-5.6 Sol?

Barely, and not on every test. The headline story is price and speed, not a jump in ability.

Here's the test-by-test comparison from Artificial Analysis, with both models at their top settings:

Test (Artificial Analysis)	GPT-6 Sol	GPT-5.6 Sol	Change
AA-Omniscience (factual reliability)	27	22	+5
Terminal-Bench 4.0 (terminal tasks)	44%	40%	+4
AutomationBench-AA (business app workflows)	62%	60%	+2
SciCode (scientific Python)	58%	57%	+1
AA-LCR v1.1 (long-context reasoning)	84%	84%	0
Humanity's Last Exam (hard questions)	48%	49%	-1
CritPt (research-level physics)	31%	32%	-1
GDP.pdf (long professional documents)	25%	27%	-2
AA-Briefcase v1.1 (multi-week knowledge work)	1483	1487	-4
GDPval-AA v2.1 (agent tasks from 44 occupations)	1487	1588	-101

The gains line up with OpenAI's claims. OpenAI says GPT-6 Sol "makes about half as many mistakes as its predecessor," based on its own factuality tests of real conversations. The independent factual-reliability score rose from 22 to 27, which points the same way.

The drop on GDPval-AA is the surprise. It's a 101-point fall on a test of real knowledge-work tasks. If your agents do that kind of work, test GPT-6 Sol before switching, rather than assuming newer is better.

What Is GPT-6 Luna Good For?

Cheap, fast work where you supply the facts. That matches OpenAI's own positioning of summarizing, extracting and answering questions.

Artificial Analysis's Luna vs Sol comparison shows where it's strong and where it isn't:

• Long documents: Luna scores 83% on long-context reasoning, almost level with Sol's 84%.
• Scientific coding: it scores 55% on SciCode, close to Sol's 58%.
• Terminal work: it scores just 13% on Terminal-Bench 4.0, against Sol's 44%. It's not an agent for command-line tasks.
• Factual recall: it scores 1 on AA-Omniscience, against 27 for Sol. That test penalizes wrong answers, so a score near zero means Luna guesses wrong about as often as it knows.

That last point needs care. OpenAI says Luna at higher effort matches GPT-5.6 Sol's factuality at about a hundredth of the cost, based on its own tests.

The independent test measures something narrower: recall without help. The practical takeaway is the same either way. Give Luna the facts through retrieval or documents, and don't rely on what it remembers.

For routing, Luna works well as a first pass. Let it handle the easy, high-volume steps, and send harder ones to Sol or a frontier model. We covered that pattern in our guide to fast classifiers for AI agents.

How Do GPT-6 Sol and Luna Compare With Other Models?

They win on cost, not capability. Here's where they sit among the models we've covered this month, on the same Artificial Analysis index:

Model and setting	Intelligence Index	Cost per task
Claude Opus 5.5 (medium)	51	$1.34
GPT-6 Astra (medium)	50	$1.54
GPT-6 Sol (max)	48	$1.06
Grok 4.7 (high)	46	$2.73
GPT-6 Sol (high)	43	$0.37
Claude Opus 5.5 (low)	42	$0.55
GPT-6 Luna (max)	37	$0.07

A few things stand out:

• GPT-6 Sol beats Grok 4.7 on value. It scores higher at its top setting for well under half the cost per task. Both charge $2 per million input tokens, though Grok's output is cheaper per token.
• Sol and Opus 5.5 overlap at the low end. Sol at high scores 43 for $0.37, while Opus 5.5 at low scores 42 for $0.55.
• For the hardest work, pay more. Opus 5.5 and Astra score higher at medium than Sol does at max.
• Luna has no close rival here at $0.07 per task, though it's also far behind on capability.

TechCrunch reports OpenAI saying Sol and Luna "substantially" outperform Anthropic's Fable and Opus models on various tasks.

The independent composite doesn't show that for overall capability, though the cost gap is large. Anthropic's new model has its own rough edges. Simon Willison found Opus 5.5 at max effort "over-thinks to breaking point," running out of output room on one test. Our Grok 4.7 analysis covers why cost per task beats price per token as a guide, and our GPT-6 Sol vs Claude Opus 5.5 comparison matches them setting by setting.

What Do OpenAI's Honesty Numbers Show?

Large improvements, by OpenAI's own measure. VentureBeat reports these results from OpenAI's testing:

• Deception rate: down to 1.3% for Sol from 10.4%, and to 2.8% for Luna from 9.5%.
• Failure to disclose problems: down to 5.4% for Sol from 77.8%, and to 30.2% for Luna from 78.3%.

These are big shifts, and they matter for agents that report on their own work. But they're OpenAI's numbers on OpenAI's tests. Luna's 30% failure-to-disclose rate is still high. For any agent that acts on its own, keep checks in place that don't rely on the model reporting its own mistakes.

GPT-6 Sol and Luna: Which Should You Use?

Match the model to the job and the budget. Here's a starting map:

• Summaries, extraction and classification at scale: GPT-6 Luna, with the facts supplied by retrieval.
• First-pass routing in agents: Luna, with harder cases sent to Sol or a frontier model.
• Everyday coding help and mid-level agent work: GPT-6 Sol at medium or high, where the cost per task is very low.
• Knowledge-work agents: test Sol carefully, given its GDPval drop, or use a stronger model.
• Terminal and computer-use agents: Sol at a minimum. Avoid Luna here.
• The hardest reasoning and agent tasks: Claude Opus 5.5 or GPT-6 Astra, which score clearly higher.

If you already use GPT-5.6 Sol or Luna, switching is mostly a model ID change at half the price. The main thing to check is whether your specific tasks are among those that improved or those that slipped.

How Should You Test GPT-6 Sol and Luna?

On your own tasks, across settings, with cost per finished task as the main number.

• Pick 50 to 200 real tasks with known good answers, split by type.
• Run Luna, Sol and your current model at low, medium and high.
• Record success, cost and time per task, and look for the cheapest setting that clears your quality bar.
• Test factual answers separately, with and without retrieval, since that's where Luna is weakest.
• Check long inputs, since Sol's context window is smaller than GPT-5.6 Sol's.
• Compare against your current baseline, so you can see whether the savings came with any drop in quality before you switch traffic.

Our guide to AI agent evaluation covers how to build and maintain a test set like this.

How Van Data Team Helps Teams Cut Model Costs Without Losing Quality

We help teams take advantage of price drops like GPT-6 Sol and Luna safely. That means routing each task to the cheapest model that handles it well, building test sets from real work, and measuring cost and quality per finished task.

With GPT-6 Sol and Luna, a 50% price cut only helps if the cheaper model still does the job. If you want to lower your AI bill without surprises, our work on AI agent evaluation and AI agent development cost is a good place to start.