Pick GPT-6 Sol for agentic coding and development work that benefits from careful multistep validation, and pick GPT-6 Luna for smaller, faster, high-volume tasks where cost per call decides everything — that is the GPT-6 Luna vs GPT-6 Sol outcome in one line, and the price gap behind it is roughly twenty to one. OpenAI announced both models on 22 September 2026, expanding the GPT-6 family that previously contained only the premium GPT-6 Astra, and the split between them is unusually clean: Sol is the balanced workhorse, Luna is the budget runner. Below is what each model is for, what they cost, which GitHub Copilot plans get them, and how to decide between them for agent building.
📺 Watch: GPT-6 Sol + Luna Just Changed AI Agents
🔥 Get the Agent OS as a free bonus: AI Profit Boardroom members get the full Agent OS zip, prompt libraries, daily tutorials and weekly live coaching calls. → Get inside
What GPT-6 Sol and GPT-6 Luna Actually Are
The clearest official description comes from GitHub's changelog entry of 22 September 2026 announcing both models in Copilot. It calls GPT-6 Sol "a balanced model for interactive and agentic coding" and "a strong all-round choice for development tasks that benefit from careful, multistep validation." GPT-6 Luna, in the same entry, is "a lightweight, cost-efficient model for smaller, faster tasks and the lowest-cost option in the GPT-6 family." Those two sentences carry the whole positioning: Sol is aimed at real engineering loops — plan, execute, check, correct — while Luna is aimed at the long tail of quick jobs where you care about speed and unit economics more than depth.
Both join GPT-6 Astra, the family's premium tier, rather than replacing it. In practice that gives OpenAI a three-rung ladder within one generation: Astra when quality is worth almost any price, Sol for everyday development, Luna for volume work. If you have been following the family since Astra shipped, the pattern mirrors what the previous generation did with its own Sol and Luna tiers — the guide to running GPT-6 in Hermes Agent covers how the family slots into an agent harness.
GPT-6 Luna vs GPT-6 Sol: Pricing Compared
Pricing is where the two models separate hardest. Per OpenAI's published rates, recorded in Simon Willison's 22 September 2026 write-up of the launch, GPT-6 Sol costs $2 per million input tokens, $0.20 per million cached input tokens, and $10 per million output tokens. GPT-6 Luna costs $0.10 input, $0.01 cached input, and $0.50 output. Luna is, in Willison's words, one of the cheapest models OpenAI has ever released — and both models land at half the price of their GPT-5.6-generation namesakes, which sat at $4/$20 for Sol and $0.20/$1.20 for Luna.
| Per 1M tokens | GPT-6 Luna | GPT-6 Sol | GPT-6 Astra |
|---|---|---|---|
| Input | $0.10 | $2 | $10 |
| Cached input | $0.01 | $0.20 | — |
| Output | $0.50 | $10 | $50 |
Sources: OpenAI pricing as recorded in Simon Willison's analysis, 22 September 2026.
Read that table like an operator, not a spectator. A task that consumes 10,000 input tokens and produces 2,000 output tokens costs about 2.1 cents on Sol and about 0.2 cents on Luna. Run that task a thousand times a month and the difference is $21 versus $2 — still small — but run an always-on agent fleet processing millions of tokens daily and Luna versus Sol becomes the difference between a rounding error and a real line item. The cached-input discount matters too: both models give a 90%-plus discount on repeated context, which rewards agent designs that hold instructions stable across calls.
If you want to build agents on models like these with the routing, prompts and workflows already solved — the AI Profit Boardroom hands you the full Agent OS zip, prompt libraries and weekly live coaching calls — step inside here. Want a personal read on your own AI and SEO setup first? Book a free SEO strategy session and get it mapped in one call.
Which Model Wins for Which Job
The honest GPT-6 Luna vs GPT-6 Sol decision comes down to how much validation your task needs. Based on the official positioning, here is the split that follows from it, best fit first:
- Agentic coding sessions — GPT-6 Sol. GitHub's description leans on "careful, multistep validation", which is precisely what long-running coding agents need: the model checks its own steps instead of sprinting to a plausible answer. If an agent writes code that ships, Sol is the default.
- High-volume pipeline steps — GPT-6 Luna. Classification, extraction, summarising, formatting, routing decisions, first-pass drafts: tasks you run thousands of times where a twenty-times price gap compounds. Luna is built to be the lowest-cost rung of the family for exactly this.
- Interactive daily driving — GPT-6 Sol. For a hands-on coding assistant in your editor, Sol's balance of quality and cost is the sensible everyday setting, with Astra held back for the genuinely hard problems.
- Draft-then-verify chains — both. A pattern worth stealing: Luna produces the cheap first pass, Sol validates and finishes. You pay Sol prices only on the fraction of tokens that survive the first cut.
If you want to see how this generation's models stack up against each other in hands-on testing rather than positioning statements, the Goldie Bench write-up covers how these brains compare across real agent tasks, and the comparison of Claude Opus 5.5 vs Claude Opus 5 covers the other big release of the same week — Anthropic's cheaper flagship landed one day before Sol and Luna, which tells you plenty about where the market is heading.
Copilot Availability and Where You Can Run Them
GitHub's changelog spells out plan access precisely: GPT-6 Sol is available to Copilot Pro+, Max, Business and Enterprise plans, while GPT-6 Luna reaches further down the ladder to Copilot Pro as well as Pro+, Max, Business and Enterprise. Both models are rolling out gradually across Visual Studio Code, Visual Studio, Copilot CLI, the Copilot cloud agent, the Copilot app, github.com, GitHub Mobile, JetBrains IDEs, Xcode and Eclipse, billed under usage-based billing. The asymmetry is deliberate — Luna serves as the affordable entry point on cheaper plans, Sol as the step-up model for paid-tier developers. Outside Copilot, both are standard API models, so the usual harnesses and routers apply; comparisons like DeepSeek Harness vs Claude Code show how much the harness around a model changes what you actually get from it.
📺 Watch: Claude Opus 5.5 + NEW GPT 6 Models!
The Price War Context Behind This Launch
Neither model exists in a vacuum. Within roughly forty-eight hours of 21–22 September 2026, xAI shipped Grok 4.7 at $2 input and $6 output, Anthropic cut Claude Opus 5.5 to $4/$20 — a 20% reduction, per Willison's same 22 September analysis — and OpenAI answered with Sol at $2/$10 and Luna at $0.10/$0.50, both half their previous-generation prices. Frontier-quality tokens are getting cheaper fast, and the practical winner is anyone running agent workloads: capability that cost real money a year ago now costs cents. It also means model choice is becoming a portfolio decision rather than a loyalty decision — the roundup of the best AI agents for coding makes the case for keeping your harness model-agnostic, and benchmark write-ups like the Kimi 2.6 benchmark breakdown show how quickly the leaderboard reshuffles underneath any single-vendor bet. Six months from now the specific numbers in this article will have moved again; the Luna-for-volume, Sol-for-validation split is the part likely to hold.
One thing this comparison deliberately does not do is crown an overall champion on benchmark scores — OpenAI's launch materials for Sol and Luna lead with positioning and price rather than a benchmark table, and honest comparison follows the evidence available. What the pricing alone establishes is enough to act on: the Agent OS pattern of routing each task type to the cheapest model that clears its quality bar was built for exactly this kind of tiered family, and the previous generation's GPT-5.5 Pro deep-dive shows how much of that older stack the 6-series pricing just made obsolete.
Quick Answers on GPT-6 Sol and Luna
Are GPT-6 Sol and Luna available right now? Yes, with a caveat: GitHub's 22 September 2026 changelog says the rollout across Copilot surfaces happens gradually, so a specific editor or plan may see the models appear over days rather than instantly. Do they replace GPT-6 Astra? No — Astra remains the premium tier of the family at $10 input and $50 output per million tokens, and Sol and Luna extend the family downwards rather than superseding it. Is there a catch in the cheap pricing? Only the usual one: both models bill under usage-based billing in Copilot, so volume still costs money — the discount rewards efficient prompt design, and the near-free cached-input rates reward agents that keep their instructions stable between calls.
GPT-6 Luna vs GPT-6 Sol: Final Verdict
Default to GPT-6 Sol for anything agentic, interactive or shipped to production; default to GPT-6 Luna for anything you run at volume where a cheap, fast pass is good enough; escalate to Astra only when a task visibly outgrows Sol. At $2/$10 versus $0.10/$0.50 per million tokens, the two models are not really competitors — they are two rungs of the same ladder, and the smart move is wiring your stack so each call lands on the right one automatically.
If you want the done-for-you routing setups, agent templates and daily tutorials that make tiered model families pay — check out the AI Profit Boardroom and get inside with the operators. Rather start with a 1-on-1? Book a free SEO strategy session and get your AI stack reviewed live.
