GPT-6 Sol vs Luna: Which to Use

For nineteen days after GPT-6 Astra launched, the GPT-6 generation was a single, expensive model. If you wanted GPT-6 capability, you paid $10 per million input tokens and $50 output, and for everything else you kept using the GPT-5.6 family. On September 22, 2026, that changed. OpenAI released GPT-6 Sol and GPT-6 Luna, filling in the mid and low tiers and cutting prices to half of what the equivalent GPT-5.6 models charged. It confirmed to VentureBeat that the new rates are permanent rather than promotional.

The timing tells you what kind of release this is. It arrived roughly ninety minutes after Anthropic shipped Claude Opus 5.5 at $4 and $20, and a day after xAI launched Grok 4.7 at $2 and $6. This is a pricing move as much as a model release, and the practical question for anyone building on the API is not whether Sol or Luna is impressive but which one each of your workloads should run on. This guide covers what the two tiers cost, what OpenAI's own data says about their performance and what it does not, how they stack up against Astra and against Anthropic, and a routing framework for using all three.

Key Takeaways

• GPT-6 Sol costs $2 input and $10 output per million tokens; GPT-6 Luna costs $0.10 and $0.50. Both launched September 22, 2026 with permanent pricing.
• The GPT-6 family is now three tiers: Luna for high-volume routine work, Sol for complex coding and professional tasks, Astra ($10/$50) for the hardest reasoning. There is no GPT-6 Terra.
• Prices are half the GPT-5.6 predecessors: Sol down from $4/$20, Luna from $0.20/$1.20, with Luna's output cut closer to 58 percent.
• On OpenAI's own charts the gains are mostly in cost per task rather than peak scores; independent analysts note GPT-5.6 Sol still scores higher on some coding benchmarks.
• Sol matches Claude Sonnet 5 exactly on every published price line, so the choice between them is fit and quality, not cost.
• The spread from Luna to Astra is 100x on input and output alike, which makes routing by difficulty the dominant cost lever.

The GPT-6 Family at a Glance

Three models, three price bands, one integration. All rates are per million tokens in US dollars, standard API pricing as of September 23, 2026.

Two details worth noting. All three tiers share the same context window, so Luna's cheapness does not come with a shorter memory, which is unusual at its price. And the family skips the middle-ground Terra tier that GPT-5.6 had; OpenAI has collapsed the lineup to three clear bands rather than four. Both new models are API-only, with no downloadable weights, and both were trained with methods OpenAI describes as similar to Astra's.

What Changed From GPT-5.6

OpenAI frames the release as a 50 percent price cut, and for Sol that is exact. For Luna it undersells the output side.

One caveat matters for anyone modeling the saving. The GPT-5.6 Sol price OpenAI compares against, $4 and $20, is itself promotional, guaranteed only through at least November 21, 2026, and the pricing page does not say what it becomes afterward. So the true long-run comparison is not quite 50 percent; it is against a predecessor whose own price was already discounted. The GPT-5.6 models remain available in the API for teams pinned to their behavior, and GPT-6 Sol has a knowledge cutoff of April 20, 2026 against GPT-5.6 Sol's February 16, 2026.

Performance: What OpenAI's Data Shows, and What It Does Not

OpenAI's launch material makes its case with cost-per-task charts plotted across five effort levels, and two headline claims stand out. Sol, it says, outperforms Claude Opus 5 on AutomationBench while costing about 9 percent as much per completed task, reaching 33.2 percent on AutomationBench 1.0.6 at its highest effort setting. Luna scores 66.6 percent on the DeepSWE v1.1 coding benchmark at roughly 93 percent lower cost than Opus 5.

Those are real and meaningful numbers, but read them for what they measure. OpenAI has deliberately shifted the framing from raw benchmark scores to cost per completed task, and on that metric the new models win decisively, because the price cut does most of the work. Read the underlying charts rather than the quoted points, as some independent analysts have, and a more nuanced picture appears: on two of the three coding and computer-use charts, GPT-5.6 Sol's best score is actually higher than GPT-6 Sol's. The improvement is largely economic, not a leap in peak capability.

That is not a criticism so much as a clarification of what you are buying. If your constraint is cost per task, Sol and Luna are substantial upgrades. If your constraint is the hardest possible score on a difficult benchmark, GPT-5.6 Sol at higher effort or GPT-6 Astra may still be the better tool. And all of these figures come from OpenAI's own evaluations run in its research environment, with competitor scores taken from public reports, so treat them as vendor claims pending independent replication rather than settled fact.

How Sol and Luna Compare to Anthropic and Everyone Else

The competitive picture is where this release gets interesting, because OpenAI has priced its two new tiers directly against Anthropic's. GPT-6 Sol at $2 and $10 lands on exactly the same rates as Claude Sonnet 5, and it matches on every line both vendors publish, including a $2.50 short cache write and a $0.20 cache read. GPT-6 Luna at $0.10 and $0.50 sits at one-tenth of Claude Haiku 4.5's $1 and $5. Our Anthropic versus OpenAI comparison covers the wider relationship between the two labs.

Notice the pattern: at both the top and the middle, OpenAI and Anthropic are now at identical prices, Astra matching Fable 5.1 and Sol matching Sonnet 5. The two labs have effectively agreed on what a frontier token and a workhorse token cost. The only unmatched band is the bottom, where Luna undercuts Haiku by an order of magnitude, and that is precisely where the pressure now sits. For the full cross-provider picture, see our guide to LLM token costs by model.

Which Tier to Use: A Routing Framework

Because the three tiers share an API and a context window, the practical question is not which model to adopt but how to route between them. The spread from Luna to Astra is 100x on both input and output, so getting this right matters more than any other cost decision you will make on the platform.

Route to Luna when the task is scoped and repetitive

Summarization, extraction, classification, tagging, routing, formatting, and routine question answering. These make up the majority of production traffic in most applications, and at $0.10 and $0.50 Luna handles them at a cost that barely registers. Its 66.6 percent DeepSWE score means it is genuinely capable, not a toy, and its full 1.05-million-token context means you can hand it large documents without a different model.

Route to Sol as the default for real work

Coding, code review, data analysis, drafting substantive documents, multi-step reasoning that is hard but not extreme. Sol is positioned as the professional workhorse and priced identically to Sonnet 5, so it is the natural default for anything Luna cannot handle reliably. For most teams, Sol should carry the bulk of the tokens that are not routine.

Route to Astra only on failure

The hardest multi-step reasoning, long-horizon agentic work, and computer-use tasks where Sol demonstrably falls short. At five times Sol's price, Astra should be the escalation path rather than the starting point. The test is empirical: run the task on Sol, measure the failure rate and what a failure costs you, and escalate only where the premium is cheaper than the errors. Our guide to GPT-6 Astra covers when the top tier earns its place.

Implemented well, this is the same routing discipline described in our LLM cost optimization guide: start each request on the cheapest tier that clears the quality bar, and escalate on failure rather than defaulting up. OpenAI's own pitch for the family, that one integration now covers every price band, is really an argument for doing exactly this.

A Worked Example

Consider a product handling 500,000 requests a month, each averaging 2,000 input tokens and 300 output tokens, which is 1 billion input tokens and 150 million output tokens.

Run entirely on Luna, that costs $100 in input and $75 in output, about $175 a month. Run entirely on Sol, it costs $2,000 and $1,500, about $3,500. Run entirely on Astra, it costs $10,000 and $7,500, about $17,500. The same volume spans a 100-fold range depending purely on which tier you chose.

Now route it. Suppose 70 percent of those requests are routine and go to Luna, 27 percent need Sol, and 3 percent escalate to Astra. The blend costs roughly $122 on Luna, $945 on Sol, and $525 on Astra, about $1,600 a month. That is less than half the all-Sol figure and less than a tenth of all-Astra, at a quality level the all-Luna configuration could not match. The routing decision, not the model release, is where the money is.

The Bottom Line

GPT-6 Sol and Luna complete OpenAI's generation and, more importantly, halve the price of using it for everyday work. Sol at $2 and $10 is a capable workhorse priced to the cent against Claude Sonnet 5; Luna at $0.10 and $0.50 is the cheapest frontier-class model on the market by a wide margin and still handles the bulk of production traffic well. Astra stays at the top for the work that genuinely needs it.

So the decision is less about Sol versus Luna than about building the routing that uses both. Send routine work to Luna, default to Sol, escalate to Astra only when Sol fails, and measure the failure rate to keep the boundaries honest. In a week where three labs cut prices in two days, the vendor rate card is the part of your AI bill you control least. Routing, caching, and output discipline are the parts you control most, and they are covered in our guide to FinOps for AI.