In September 2026, OpenAI finished assembling its GPT-6 family in the space of four weeks. Astra arrived first as the flagship, then Sol and Luna landed on September 22 at half the price of the GPT-5.6 models they replace, and a refreshed GPT-6.1 Sol followed on September 29. Three models, one lineup, and prices that span from $0.10 to $50 per million output tokens.
That spread is exactly the point. OpenAI now fields a model for enterprise-grade reasoning, one for everyday professional work, and one for high-volume tasks where cost matters more than brilliance. Here is how the three compare, what the benchmarks actually say, and how to pick the right one for your workload.
Meet the three models
GPT-6 Astra: the flagship
Astra has been OpenAI's top model since September 3, 2026. It leads on the hardest problems: complex computer-use tasks, agentic coding, and cybersecurity work. At $10 per million input tokens and $50 per million output tokens, it is also the most expensive model OpenAI currently sells — 50x the output price of Luna. Model ID: gpt-6-astra.
Astra earns that price when failure is expensive. Multi-hour coding agents, security analysis, and tasks where a single mistake costs more than a few dollars of compute are its natural habitat.
GPT-6 Sol: the everyday workhorse
Sol is the model OpenAI intends most developers to reach for. Released September 22, 2026 at $2 per million input and $10 per million output tokens, it undercuts its GPT-5.6 Sol predecessor by exactly 50% on both sides. The follow-up GPT-6.1 Sol (September 29) keeps those same rates with an even cheaper cached-input price of $0.10 per million tokens. Model ID: gpt-6-sol.
Sol is aimed at complex coding, multi-step reasoning, and professional analysis — the bulk of real developer workloads. It is also the efficiency play: OpenAI reports that GPT-6 caching now yields higher hit rates by default, and that changing reasoning effort or tool availability no longer breaks the cache.
GPT-6 Luna: the budget speedster
Luna is the cheapest current OpenAI model at $0.10 per million input and $0.50 per million output tokens, launched the same day as Sol. It replaces GPT-4o as the default choice for fast, high-volume work: summarising, extracting information, quick questions, classification, and simple automation. Model ID: gpt-6-luna.
Luna is a price cut first and a capability story second. Independent testing found small regressions against GPT-5.6 Luna on some coding evaluations even as the price fell roughly 60% per task. For text-heavy clerical work, that trade-off is usually invisible. For code, it matters.
Shared technical foundation
All three models share the same core specs: a context window of about 1.05 million tokens, up to 128,000 tokens of output in a single response, and text plus image input. The difference between them is not what they can technically handle — it is how much reasoning effort and cost goes into each answer.
One caveat worth knowing: when Sol and Luna launched, they were not yet available in standard ChatGPT chat. They shipped first through the API, ChatGPT Work, and Codex. Check your plan if you expect them in the regular chat interface.
Pricing side by side
Standard API rates per million tokens:
| Model | Input | Cached input | Output |
|---|---|---|---|
| GPT-6 Astra | $10.00 | $1.00 | $50.00 |
| GPT-6.1 Sol | $2.00 | $0.10 | $10.00 |
| GPT-6 Sol | $2.00 | $0.20 | $10.00 |
| GPT-6 Luna | $0.10 | $0.01 | $0.50 |
Two extras change the math for heavy users. The Batch API cuts another 50% off those rates — GPT-6 Sol batch output drops to $5 per million tokens, Luna batch output to $0.25. And long-context pricing kicks in when a request carries more than 272,000 input tokens: the higher rate applies to the entire request, not just the tokens above the threshold. At long context, Sol moves to $4/$15 and Luna to $0.20/$0.75 per million, while Astra jumps to $20/$75.
For repository agents and long-running conversations, context design — cache hit rates, tool definitions, compaction — can save more money than switching reasoning levels. Sol's improved caching is aimed directly at that.
What the benchmarks say
OpenAI's own launch comparison puts GPT-6 Sol on the cost-efficiency frontier against both prior OpenAI models and Anthropic's flagship models:
- AutomationBench (business workflows, highest effort): Sol scores 33.2% at $0.27 per task, beating Astra at 30.3% while costing 3.9x less per task.
- Agents' Last Exam (maximum effort): Sol reaches 56.4%, above Claude Opus 5's best score, at 60% lower cost per task.
- FrontierCode v1.1: Sol matches Claude Fable 5.1 at its highest reasoning effort, at what OpenAI describes as much lower cost.
- DeepSWE v1.1 (maximum effort): Sol hits 68.8%, within 1.1 points of Fable 5's best score, at roughly 80% lower cost per task.
Independent evaluators published a more skeptical read that still supports the core story. Artificial Analysis found GPT-6 Sol improved 2 points over GPT-5.6 Sol on their Coding Agent Index — real gains on Terminal-Bench 4.0 (43% vs 37%) and SWE-Atlas-QnA (58% vs 54%) at roughly half the cost per task. Sol is a genuine, if modest, capability upgrade.
Luna's independent results are mixed. The same index scored GPT-6 Luna 2 points *lower* than GPT-5.6 Luna, with declines on SWE-Atlas-QnA (44% vs 49%) and DeepSWE v1.1 (64% vs 66%), even as cost per task fell about 60%. The direct implication: Sol is an upgrade at half the price; Luna is mostly a price cut with a small quality trade-off attached. Keep Luna away from code you care about.
On factual accuracy, OpenAI claims GPT-6 Sol makes about half as many factual errors as GPT-5.6 Sol on its internal factuality evaluation, approaching Astra-level reliability. Treat vendor benchmarks as reported claims rather than independent verification, but the direction — a cheaper model that also errs less — is the pairing developers have been waiting for.
Which model should you choose?
Choose Astra when the cost of being wrong exceeds the cost of tokens. Security audits, complex multi-step coding agents on critical systems, and high-stakes analysis where a retry loop costs more than the model itself. Astra should earn its place through higher success rates, fewer retries, and better outcomes on tasks where failure is expensive.
Choose Sol for nearly everything else professional. Agentic coding, business workflow automation, deep research, and analysis. It matches or beats the flagship on several business benchmarks at a fraction of the cost, and the improved prompt caching compounds the savings. If you are migrating from GPT-5.6 Sol or GPT-4o, Sol is the natural new default.
Choose Luna for high-volume, latency-sensitive work. Summaries, extraction, classification, chatbots, content pipelines — anywhere you run thousands of calls and the task is well within a small model's reach. At $0.10 per million input tokens, it sits below every frontier-class model on the market. Just do not expect it to write your software.
A practical pattern that is emerging: route with a router. Send the first attempt to Luna, escalate unclear cases to Sol, and reserve Astra for the few decisions that justify it. The three-tier family was clearly designed for exactly that strategy — a developer no longer leaves the OpenAI ecosystem regardless of budget.