OpenAI just expanded the GPT-6 family. On September 22, 2026, the company quietly rolled out GPT-6 Sol and GPT-6 Luna, two new models that borrow the best ideas from the flagship GPT-6 Astra and squeeze them into faster, cheaper packages. There was no keynote and no countdown clock. The models simply appeared across ChatGPT, Codex, and the API.
The timing matters. GPT-6 Astra landed on September 3 and immediately took the crown as OpenAI's most capable system. Less than three weeks later, Sol and Luna arrived to fill in the lower tiers. The message is clear: OpenAI doesn't want GPT-6 to be a single premium product. It wants a lineup that covers everything from brutal coding problems to summarizing a few hundred pages of invoices before lunch.
What Are GPT-6 Sol and GPT-6 Luna?
Think of Astra as the model you call when the task is genuinely hard. Sol and Luna handle the volume.
GPT-6 Sol is built for demanding professional work: complex code refactoring, long-horizon agent workflows, and multi-step reasoning chains that don't forgive sloppy logic. OpenAI describes it as the option for tasks that need deep thinking and careful follow-through.
GPT-6 Luna is the speed specialist. It targets high-throughput jobs where the goal is obvious and the volume is huge. Extracting invoice data from hundreds of PDFs, translating chat logs in real time, or answering a constant stream of routine questions. Luna isn't trying to write Shakespeare. It's trying to do a narrow job fast and cheap, thousands of times over.
Both models inherit Astra's progress on professional work, factual accuracy, coding, computer use, and alignment. OpenAI's own framing is simple: build with Sol, scale with Luna.
GPT-6 Release Date and Availability
Here's where you can actually use them, starting today.
| Plan | ChatGPT Work | Codex | API | Standard ChatGPT chat |
|---|---|---|---|---|
| Plus | Yes | Yes | Yes | Not yet |
| Pro | Yes | Yes | Yes | Not yet |
| Business | Yes | Yes | Yes | Not yet |
| Enterprise | Yes | Yes | Yes | Not yet |
| Edu | Yes | Yes | Yes | Not yet |
| Free | Luna only (desktop app) | No | No | Not yet |
| Go | Luna only (desktop app) | No | No | Not yet |
Sol and Luna are now live for Plus, Pro, Business, Enterprise, and Edu subscribers across ChatGPT Work and Codex, and through the API. Free and Go users get GPT-6 Luna on the desktop app. One catch: neither model has reached the standard ChatGPT chat interface yet. OpenAI says that rollout is coming.
If you're on GPT-5.5, pay attention. OpenAI has confirmed that GPT-5.5 will retire from ChatGPT, ChatGPT Work, and Codex on October 14, 2026. The new model tier is meant to absorb that traffic.
GPT-6 New Features: Price, Speed, and Fewer Mistakes
The headline number is price. Both models come in at 50% below GPT-5.6's promotional pricing, and OpenAI says the cut is permanent, not a limited-time promotion.
| Model | Input (per 1M tokens) | Output (per 1M tokens) |
|---|---|---|
| GPT-6 Sol | $2 | $10 |
| GPT-6 Luna | $0.10 | $0.50 |
| GPT-6 Sol (GPT-5.6 promo) | $4 | $20 |
| GPT-6 Luna (GPT-5.6 promo) | $0.20 | $1.20 |
OpenAI credits better caching and inference efficiency for the drop, and says it's passing the savings straight to users rather than pocketing the margin.
Then there's accuracy. Based on OpenAI's internal testing on de-identified real-world conversations where users flagged mistakes, GPT-6 Sol's error rate is roughly half that of its predecessor. That gets it close to Astra-level reliability while costing far less. Luna also shows a meaningful jump in factual reliability, again at a lower price.
The company claims Sol and Luna beat Anthropic's top models on task performance. It's worth reading those claims with some caution. On OpenAI's own AutomationBench cross-application business workflow test, GPT-6 Sol at xhigh effort beat Claude Opus 5 at just 9% of Opus 5's per-task cost. On the agent-focused Last Exam, GPT-6 Sol hit 56.4% at max effort, above Claude Opus 5's best score, while cutting per-task cost by 60%.
Coding gets specific numbers. On DeepSWE v1.1, GPT-6 Sol reaches 68.8% at max effort, just 1.1 points behind Claude Fable 5's top score of 69.9%, at roughly 80% lower cost per task. GPT-6 Luna lands at 66.6% while running about 93% to 96% cheaper per task than Opus 5 and Fable 5 at medium effort.
For computer use, OpenAI is refreshingly blunt: Astra is still the strongest model for this. Sol and Luna just do the same work at a lower cost. On OSWorld 2.0 offline, GPT-6 Sol at xhigh effort posts 60.5%, close to Claude Opus 5's 60.3% at medium effort, at around 80% less cost per task.
There's a softer update too. OpenAI carried over Astra's improved communication style. In technical and coding conversations, the new models aim to be clearer, use less jargon, avoid odd phrasing, trim low-value detail, and shorten answers without losing substance.
Why OpenAI Is Splitting the GPT-6 Lineup
This isn't a random product drop. It's a pattern that's been building across the AI industry.
Frontier labs increasingly ship a flagship to push the ceiling and cheaper, faster variants to push adoption. Astra handles the hardest, highest-stakes work. Sol and Luna bring pieces of that capability into everyday, high-frequency workflows. The result is that advanced model ability, once reserved for rare and expensive calls, starts showing up in routine coding, agent tasks, and enterprise automation at a fraction of the unit cost.
For developers and businesses, the practical effect is that the calculus changes. Tasks that never justified a premium model call suddenly become affordable to run at scale. That's a bigger shift than another benchmark record.
The Competition Never Slows Down
Sol and Luna didn't launch into a quiet market. Anthropic shipped an Opus 5.5 update roughly 90 minutes before OpenAI's announcement, which tells you how tight this race has become. OpenAI's repeated comparisons to Claude Fable and Opus read as a direct response.
OpenAI researcher Noam Brown has described a "multiplier effect" in this generation between pretraining and chain-of-thought reinforcement learning. That's his explanation for why GPT-6 models show steep gains on complex tasks. Whether that compounds into durable progress or just better benchmark numbers is the open question.
What to Watch Next
A few things are worth tracking over the coming weeks.
First, the standard ChatGPT chat rollout. Sol and Luna living only in ChatGPT Work, Codex, and the API means most casual users haven't touched them yet. When they hit the main chat interface, adoption could jump fast.
Second, the GPT-5.5 retirement on October 14. That deadline will push a lot of users onto the new tier whether they planned for it or not.
Third, OpenAI DevDay 2026, set for September 29. Given how aggressive the release pace has been, expect more.
Final Thoughts
GPT-6 Sol and Luna aren't trying to be the smartest models on the planet. That job belongs to Astra. Instead, they take a slice of frontier intelligence and make it cheap, fast, and easy to run at volume. The 50% price cut and the claim of halved error rates are the real story here. If those hold up in everyday use, the bigger winner might not be the benchmark leader. It might be the developer who can suddenly afford to run a capable model on every request instead of saving it for the important ones.
If you want to try GPT-6 Luna without a paid plan, the desktop app is the place to start. For everything heavier, Sol is the model to reach for.