The ChatGPT-6 Sol vs Luna choice is the most practical AI decision you’ll make this month. On September 22, 2026, OpenAI released two GPT-6 models at once. Sol is the capable workhorse. Luna is the very cheap one for high volume. OpenAI also cut API prices in half compared with the GPT-5.6 generation. If you run chat workflows, coding agents, or anything billed by the token, your costs and your default model just changed. Below: what shipped, what the benchmarks show, and which model to use for each job.
What’s new in ChatGPT-6 Sol vs Luna
The official names are GPT-6 Sol and GPT-6 Luna, and the API model IDs are gpt-6-sol and gpt-6-luna. Both sit below GPT-6 Astra, the flagship OpenAI previewed on September 3. OpenAI built Astra for maximum capability and built Sol and Luna for cost efficiency. All three GPT-6 models share a 1.05 million token context window and up to 128K output tokens. They accept text and image input and return text only.
Pricing
Pricing is the headline. Sol costs $2 per million input tokens and $10 per million output tokens, half the price of GPT-5.6 Sol. Luna costs $0.10 input and $0.50 output, 50% cheaper on input and about 58% cheaper on output than GPT-5.6 Luna. OpenAI says these prices are permanent, not introductory. Cached input gets a 90% discount, so Sol cache reads cost $0.20 per million tokens and Luna’s cost $0.01. One catch: requests over 272K input tokens are billed at 2x the input rate and 1.5x the output rate.
Reasoning, accuracy, and availability
Both models offer six reasoning effort levels: none, low, medium, high, xhigh, and max. The default is medium. OpenAI also claims large gains in factual accuracy. By its internal numbers, Sol makes roughly half as many mistakes as GPT-5.6 Sol, and Luna matches GPT-5.6 Sol’s factuality at about one-hundredth of the task cost.
OpenAI is rolling both models out gradually in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu users. Free and Go users get Luna in the desktop app. In the API, both models work with the Responses and Chat Completions endpoints, including the Batch and Flex options.
Why it matters
- Luna is now a serious default for bulk work. At $0.10 input and $0.50 output, classification, routing, extraction, and summarization at scale cost almost nothing. For many pipelines, the cost case for self-hosting small open-weight models has mostly disappeared.
- Sol competes on cost per task, not peak score. On AutomationBench, OpenAI’s multi-tool business workflow test, Sol at xhigh effort scores 33.2% for about $0.27 per task. OpenAI’s chart puts Claude Opus 5 at 26.9% for roughly nine times the cost per task. That gap is the whole pitch.
- Sol does not beat its predecessor at everything. Reported benchmarks show Sol at 68.8% on DeepSWE 1.1, below GPT-5.6 Sol’s 72.7%. If you push GPT-5.6 Sol hard on complex coding, test before switching.
- Luna trails Sol only slightly on coding. Luna scores 66.6% on DeepSWE, 2.2 points behind Sol, for a fraction of the cost per task. That makes Luna a strong choice for coding sub-agents and parallel workers.
- Long-context work costs more than the headline price. Above 272K input tokens, the surcharge shrinks Sol’s price advantage. If you routinely send whole codebases or large document sets, calculate costs at the surcharged rate.
- The comparisons are OpenAI’s own. Anthropic released Claude Opus 5.5 the same week, and it does not appear on OpenAI’s charts; OpenAI compared against Opus 5 instead. Don’t treat vendor charts as the final word.
How to use ChatGPT-6 Sol and Luna today
-
Pick the model in ChatGPT. On Plus, Pro, Business, Enterprise, or Edu, open the model picker in ChatGPT Work or Codex and choose GPT-6 Sol or GPT-6 Luna. The rollout is gradual, so if you don’t see them yet, check again in a day or two. On the Free or Go plan, use the desktop app to get Luna.
-
Make your first API call with Luna. Start with the cheap model and move up only if quality falls short. A minimal Responses API call:
curl https://api.openai.com/v1/responses \ -H "Authorization: Bearer $OPENAI_API_KEY" \ -H "Content-Type: application/json" \ -d '{ "model": "gpt-6-luna", "reasoning": { "effort": "low" }, "input": "Classify this support ticket as billing, bug, or feature request: The export button crashes the app." }' -
Use Sol for agent and automation work. Sol posted its best AutomationBench result at xhigh effort. Set it explicitly:
from openai import OpenAI client = OpenAI() resp = client.responses.create( model="gpt-6-sol", reasoning={"effort": "xhigh"}, input="Review this migration plan and list every step that could cause downtime, with a fix for each." ) print(resp.output_text) -
Route by task, not by habit. The biggest savings come from sending each request to the cheapest model that can handle it. A simple router is enough to start:
def pick_model(task_type: str) -> tuple[str, str]: routes = { "classify": ("gpt-6-luna", "low"), "summarize": ("gpt-6-luna", "medium"), "subagent": ("gpt-6-luna", "max"), "agent": ("gpt-6-sol", "xhigh"), "review": ("gpt-6-sol", "high"), } return routes.get(task_type, ("gpt-6-sol", "medium")) model, effort = pick_model("summarize") resp = client.responses.create( model=model, reasoning={"effort": effort}, input=document_text, ) -
Structure prompts for the cache discount. Cached input is 90% cheaper. Put stable content such as system instructions, tool definitions, and reference docs at the start of the prompt, and put the variable part at the end. For a repeated system prompt, this one change often saves more than switching models.
SYSTEM (stable, cached): You are a support triage assistant for Acme. Categories: billing, bug, feature, account. Respond with JSON: {"category": ..., "urgency": 1-5}. [... full policy doc ...] USER (variable, not cached): Ticket #48213: "I was charged twice this month." -
Benchmark on your own data before migrating. Pull 50 to 100 real prompts from your logs and run them through
gpt-6-luna,gpt-6-sol, and your current model. Compare quality, latency, and total cost. Given Sol’s reported drop on DeepSWE, this matters most for heavy coding workloads.
How ChatGPT-6 Sol vs Luna compares
The obvious rival is Claude Opus 5.5, with GPT-6 Astra above both OpenAI models. Prices are list API rates per million tokens. Benchmark figures come from OpenAI’s launch materials and early third-party reporting, so treat them as directional.
| Model | Input / Output (per 1M) | Cached input | Reported highlights | Best for |
|---|---|---|---|---|
| GPT-6 Luna | $0.10 / $0.50 | $0.01 | 66.6% DeepSWE 1.1 (~$0.22/task) | High-volume tasks, coding sub-agents, routing |
| GPT-6 Sol | $2 / $10 | $0.20 | 33.2% AutomationBench (xhigh, ~$0.27/task); 68.8% DeepSWE; 56.4% Agents’ Last Exam | Default agents, business automation, general chat |
| GPT-6 Astra | $10 / $50 | $1.00 | OpenAI’s flagship; leads OpenAI’s lineup on computer use and SWE | Hardest agentic and computer-use work |
| Claude Opus 5.5 | $4 / $20 | $0.40 | Reportedly 40.0% AutomationBench; not on OpenAI’s charts | Top-end coding, long sessions, complex reasoning |
Our take: at $2/$10, Sol is the new price-performance default for general agent work. Luna is the obvious pick when volume matters more than peak quality. If you are paying for raw capability on hard coding or multi-step automation, Opus 5.5 still looks stronger on the available numbers: its reported 40.0% on AutomationBench beats Sol’s 33.2%, though it costs twice as much per token. Run a head-to-head test on your own workload before committing.
What’s next
First, watch for independent benchmarks. OpenAI’s launch charts compared Sol and Luna against Claude Opus 5 and Fable 5, not the newer Opus 5.5 or Fable 5.1. OpenAI also skipped same-harness comparisons against Gemini 3.8 Flash and Grok 4.7. Third-party evaluations over the next few weeks will either confirm or narrow Sol’s cost-per-task lead.
Second, watch the GPT-5.6 timeline. OpenAI says GPT-5.6 Sol and Luna stay available at current promotional pricing through at least November 21, 2026, and hasn’t said what happens after that. If your production stack depends on GPT-5.6, especially for coding, where it still beats GPT-6 Sol on DeepSWE, run your migration tests now rather than in late November.
Third, expect a price response. A 50% cut on the mid-tier and a Luna model at $0.10 input will pressure Anthropic, Google, and xAI. Opus 5.5 already launched 20% below Opus 5. OpenAI’s 2026 lineup of Astra, Sol, and Luna also confirms the industry’s shift to model families, where choosing the right tier matters as much as choosing the right vendor.
Frequently Asked Questions
What is the difference between ChatGPT-6 Sol and Luna?
Sol is the more capable mid-tier model, aimed at agents, automation, and general-purpose work. Luna is the low-cost model for high-volume tasks. Luna costs 20 times less per token than Sol yet comes close on coding benchmarks: 66.6% vs 68.8% on DeepSWE 1.1.
How much does ChatGPT-6 cost in the API?
GPT-6 Sol costs $2 per million input tokens and $10 per million output tokens. GPT-6 Luna costs $0.10 input and $0.50 output. Cached input is 90% off. Requests over 272K input tokens are billed at 2x the input rate and 1.5x the output rate. GPT-6 Astra, the flagship, costs $10/$50.
Can I use GPT-6 Luna for free?
Yes. Free and Go ChatGPT users get GPT-6 Luna in the desktop app. Sol requires a paid plan: Plus, Pro, Business, Enterprise, or Edu.
Is GPT-6 Sol better than GPT-5.6 Sol?
It is much cheaper and more factually accurate, but not better on every benchmark. Reported results show GPT-6 Sol slightly behind GPT-5.6 Sol on DeepSWE, a coding benchmark, and on OSWorld 2.0, a computer-use benchmark. The main gain is cost per task, not peak capability.
Should I use ChatGPT-6 Sol or Claude Opus 5.5 for coding?
For the hardest coding work, Opus 5.5 or GPT-6 Astra is the safer bet on current data. For high-volume coding agents where cost matters, Sol, or Luna at max effort for sub-agents, delivers far more work per dollar. Test both on your own repository before deciding.
What reasoning effort should I use?
Start at the default, medium. Use low or none for classification and extraction. Use xhigh or max for multi-step agent tasks, where Sol posted its best AutomationBench and DeepSWE scores. Higher effort consumes more output tokens, so match the effort level to the task’s difficulty.
Go deeper than this article
This article covers the essentials. Our premium eguide “ChatGPT vs Claude vs Gemini” gives you the full step-by-step playbook — prompts, workflows, and copy-paste recipes you can put to work today.
