The short version
GPT-6 Astra, released earlier this month, is OpenAI's most capable and aligned model — but it's overkill for routine work. Sol and Luna are the mid-tier answer: they reuse Astra's training approach across professional work, factuality, coding, computer use, and alignment, then pair it with a big price cut[1]. OpenAI frames the release as "advancing the frontier on cost efficiency" — in practice, it means most of the GPT-6 gains without paying for the full flagship.
Price: 50% cheaper than GPT-5.6
The headline is pricing. OpenAI cut API prices for both models in half compared with GPT-5.6 promotional pricing[1]:
| Model | Input (per 1M tokens) | Output (per 1M tokens) | Change |
|---|---|---|---|
| GPT-5.6 Sol → GPT-6 Sol | $4 → $2 | $20 → $10 | 50% cheaper |
| GPT-5.6 Luna → GPT-6 Luna | $0.20 → $0.10 | $1.20 → $0.50 | 50% cheaper |
GPT-6 API pricing from OpenAI's announcement, September 2026[1].
Those savings stack with new caching defaults: OpenAI improved prompt caching for GPT-6 to deliver higher cache hit rates and a 90% discount on cached input-token reads, which matters most for agents and long conversations that re-feed the same context[1]. GitHub reports the same caching work has cut the share of prompt tokens needing fresh processing by more than half across billions of requests to Copilot[1].
Benchmarks vs the field
On OpenAI's benchmark tables, Sol is positioned to beat similarly priced competitors — especially Anthropic's Claude Opus 5 — on cost-adjusted results[1]:
| Benchmark | GPT-6 Sol | Claude Opus 5 | Claude Fable 5.1 |
|---|---|---|---|
| AutomationBench 1.0.6 | 33.2% (xhigh) | 26.9% (max) | 31.4% (max, w/ Opus 5 fallback) |
| Agents' Last Exam V1 | 56.4% (max) | below Sol | — |
| DeepSWE v1.1 | 68.8% (max) | comparable to Luna | 69.9% (Fable 5, xhigh) |
| OSWorld 2.0 offline | 60.5% (xhigh) | 60.3% (medium) | — |
Selected results from OpenAI's GPT-6 Sol and Luna announcement[1]. Cost-adjusted details in the text below.
The nuance is cost. On AutomationBench, GPT-6 Sol at xhigh effort scores 33.2% at $0.27 per task — while Claude Opus 5 at max effort scores 26.9% at 11.1× that cost per task[1]. On DeepSWE v1.1, Sol at max effort hits 68.8%, within 1.1 points of Claude Fable 5's top score at roughly 80% lower cost per task[1]. Luna at max effort scores 66.6% at 93% less than Opus 5 and 96% less than Fable 5 per task[1].
Availability: where to use them
GPT-6 Sol and Luna are available starting today in ChatGPT Work and Codex for all Plus, Pro, Business, Enterprise, and Edu users[1]. Free and Go users get access to GPT-6 Luna in the desktop app. They're not yet in the standard ChatGPT Chat tab, and OpenAI is rolling them out gradually throughout the day to keep service stable — so if you don't see them right away, check back later[1]. In the API they appear as gpt-6-sol and gpt-6-luna[1].
Caching: built for agents
Alongside the price cut, OpenAI added tooling to help developers squeeze more out of prompt caching: a Prompt Caching Dashboard to see how much input is cached, a diagnostics tool that explains missed cache opportunities, and explicit breakpoints to control where cached prefixes end[1]. You can now also change reasoning effort and enable or disable tools mid-conversation without breaking the cache[1] — directly relevant if you build long-running agents.
Factuality and alignment
On OpenAI's internal factuality evaluation — built from real conversations where users flagged mistakes — GPT-6 Sol makes about half as many errors as its predecessor, approaching Astra-level reliability at much lower cost[1]. Both models also show lower rates of misleading claims about their coding work in alignment evaluations, inheriting Astra's alignment approach[1].
What it means for small business & creators
- Cheaper automation. If you run AI agents or API workflows on GPT-5.6, Sol and Luna cut the token bill in half before caching discounts[1].
- No plan change needed. Paid ChatGPT users get them in ChatGPT Work and Codex as model options; Free/Go users can try Luna in the desktop app[1].
- Better writing quality. Both inherit Astra's clearer, less jargon-heavy communication style — noticeable in technical and coding conversations[1].
The caveats
These are mid-tier models — GPT-6 Astra still wins on raw capability, so demanding research or the hardest computer-use tasks should stay on Astra[1]. The benchmark comparisons come from OpenAI's own evaluation environment, and competitor scores were pulled from public reports, so treat exact gaps as directional rather than gospel[1]. And they're not in the standard Chat tab yet, which will confuse anyone expecting them next to GPT-5.6 in the normal picker[1].
Bottom line
Verdict: If you run GPT-5.6 Sol or Luna through the API, switch — same job, half the price, plus a caching discount. If you're on ChatGPT Plus or higher and use Work or Codex, it's a free quality bump with no downside. Only hold off if you live in the standard Chat tab and don't want to change your workflow.
Frequently asked questions
How much do GPT-6 Sol and Luna cost?
GPT-6 Sol costs $2 per 1M input tokens and $10 per 1M output tokens; GPT-6 Luna costs $0.10 input and $0.50 output — both 50% cheaper than GPT-5.6 promotional pricing[1].
Are they available in ChatGPT Chat?
Not yet. They're in ChatGPT Work and Codex for paid users, and GPT-6 Luna is in the desktop app for Free and Go users[1].
How do they compare to GPT-6 Astra?
Astra remains the most capable model. Sol and Luna are cheaper mid-tier options that inherit Astra's training approach but trade some depth for lower cost[1].
Sources
- OpenAI — "Introducing GPT-6 Sol and Luna" — openai.com
- OpenAI — "GPT-6 Astra" — openai.com
- OpenAI Deployment Safety — GPT-6 Astra system card — deploymentsafety.openai.com