GPT-6 Sol and Luna Explained: Pricing, Tiers, and Uses

OpenAI just widened the GPT-6 family, and the naming makes it easy to get lost. If you searched for GPT-6 Sol and Luna to work out what they actually are and whether they matter for what you are building, this is the short version: two cheaper, faster models that bring most of the flagship's ability to everyday work.

They sit below GPT-6 Astra, they cost far less than the previous generation, and they are already usable in OpenAI products and the API. The rest of this guide covers what each one does, the pricing, the honest benchmark picture, and how to pick a tier without a computer science degree.

GPT-6 Sol and Luna bring GPT-6 intelligence to cheaper tiers

GPT-6 Sol and Luna are two new models OpenAI released on September 22, 2026. They sit below GPT-6 Astra, the flagship that launched earlier the same month. OpenAI trained them with the same methods as Astra, then tuned them for lower cost and faster responses.

1. GPT-6 Sol is the balanced coding tier

Sol is the model to reach for on real development work. It is built for interactive and agentic coding, which means multistep tasks like building a feature, reviewing code, debugging, or analyzing data. Its API name is gpt-6-sol.

2. GPT-6 Luna is the low-cost, high-volume tier

Luna is the cheapest model in the GPT-6 family. It is built for focused, high-volume tasks such as summarizing documents, pulling out specific information, or answering quick questions. Its API name is gpt-6-luna.

3. Both tiers share Astra's context window and controls

Sol and Luna share the same 1,050,000-token context window and the same reasoning-effort controls, which range from none through low, medium, high, and xhigh up to max. A useful detail for cost control: both keep the "none" setting, so you can switch reasoning off entirely for simple, latency-sensitive calls. Astra cannot do that.

GPT-6 Sol is not the same as GPT-5.6 Sol

This is the trap most people hit, so it is worth clearing up first. OpenAI reused tier names across generations, which means there are now two different models called Sol and two called Luna.

The older GPT-5.6 family had Sol, Terra, and Luna. The new GPT-6 family has Astra, Sol, and Luna. There is no GPT-6 Terra, and OpenAI has not announced one. So when a pricing table or benchmark just says "Sol," check the generation before you trust the number. The full GPT-5.6 Sol comparison breaks the difference down further.

The quick way to tell them apart is price. GPT-6 Sol costs exactly half of the current GPT-5.6 Sol rate. In practice, GPT-6 Sol is the upgrade path for anyone who was running GPT-5.6 Sol and does not need to pay for Astra.

GPT-6 Sol and Luna cost 50% less than the models they replace

The headline is the price cut. Both models launched at roughly half the cost of the GPT-5.6 versions' promotional pricing, which OpenAI credits to improvements in caching and inference.

GPT-6 standard short-context API pricing, per 1 million tokens (as of September 2026)

For anyone building on the API, the caching changes may matter more than the sticker price. Cached input reads now get a 90% discount, and the cache holds even when you change reasoning effort or turn tools on and off between turns. Writing to the cache carries a small premium over the standard input rate, so the savings come from reading reused context across a session. The full GPT-6 Sol pricing breakdown covers the cost math in detail.

One catch to plan for: prompts above 272,000 input tokens are billed at higher rates for the entire request, not just the overage. Above that line, the input rate roughly doubles. If you routinely load huge document sets, the fix is usually to retrieve the relevant slice rather than send everything.

What the benchmarks say about GPT-6 Sol and Luna

The benchmark story has two sides, and they point in slightly different directions. Reading both is the honest way to judge these models.

1. What OpenAI reported

These are vendor-reported scores from OpenAI's own testing. On AutomationBench, a test of business workflows across apps, Sol at xhigh effort scored 33.2% at $0.27 per task, ahead of Claude Opus 5's 26.9% at max effort and at a fraction of the cost per task. On DeepSWE v1.1, a software-engineering test, Sol at max effort reached 68.8%, within 1.1 points of Claude Fable 5 at roughly 80% lower cost per task. On computer-use tasks, Sol landed near Claude Opus 5 while Astra stayed well ahead.

OpenAI also reported that, in its internal testing, Sol makes about half as many factual mistakes as GPT-5.6 Sol. You can read the figures in full in OpenAI's launch announcement. The GPT-6 Sol benchmarks page collects the same scores with the effort levels labeled.

One point OpenAI is upfront about: on some raw scores, Sol actually sits slightly below GPT-5.6 Sol. The pitch is value, not peak capability. You get similar results to far more expensive models for much less money per task.

2. What independent testing found

Independent results from Artificial Analysis tell a more measured story: a mix of small gains and small regressions, with cost as the one clear win. On its Intelligence Index, Sol landed roughly level with the previous generation. Its Coding Agent Index rose 2 points to 57, while Luna's fell 2 points to 41. The standout was cost, with Sol's spend per task on the index dropping from $1.99 to $1.06.

The hallucination numbers come with a caveat. Sol's hallucination rate fell from 92% to 60%, but part of that is because it now answers fewer questions, attempting 83% versus 99% before. Luna also regressed on some knowledge-work tasks, and its coding score slipped, so it is a weaker choice for hard, multi-file programming. The Luna benchmarks page has the independent figures side by side.

Sol, Luna, or Astra: which tier fits your work

Think of the three tiers as a ladder from cheap and fast to powerful and pricey. The right pick depends on the job, not on always reaching for the strongest model.

How the GPT-6 tiers compare for common builder jobs

For most apps, Sol is the sensible default. It carries the bulk of the work at a fifth of Astra's listed token rates, and you can dial reasoning effort up for hard steps or down for easy ones. Luna earns its place on the routine, repetitive tasks that run constantly in the background.

It helps to know the competitive backdrop too. Anthropic shipped Claude Opus 5.5 about 90 minutes before this release, and Sol lands at half of Opus 5.5's price. The Opus 5.5 comparison weighs the two head to head.

Where you can use GPT-6 Sol and Luna

Access is already fairly wide, though the rollout is gradual. Both models are available in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu accounts, and both are in the OpenAI API as gpt-6-sol and gpt-6-luna.

Free and Go users get GPT-6 Luna in the desktop app. Sol stays on the paid plans. Neither model is in the main ChatGPT chat surface yet.

Both also landed in GitHub Copilot. Sol is available on Copilot Pro+, Max, Business, and Enterprise plans, and Luna reaches Copilot Pro as well.

Start building with GPT-6 Sol today

GPT-6 Sol and Luna do not change what OpenAI's models can do so much as change what they cost. Sol brings near-flagship coding and agent work down to an everyday price, and Luna makes it cheap to run simple tasks at scale. For most builders, that turns "which model can I afford" into "which tier fits this job," which is a much better question to be asking.

If you want to build with GPT-6 Sol without wiring up API keys, you can do it on Emergent. GPT-6 Sol is available on Emergent through the Universal Key, a single credential that gives you GPT, Claude, and Gemini models with one unified bill, so you can describe the app you want and let the agent handle the setup.

Start Building on Emergent and put GPT-6 Sol to work on your next project.