First: there are two generations named Sol and Luna
Since September 22, 2026 there are two generations of models called “Sol” and “Luna”. Search results and comparison articles mix them up, so start here.
| Generation | Model | Price (in/out) | Our page |
|---|---|---|---|
| Older (GPT-5.6 era) | GPT-5.6 Sol | $4 / $20 | GPT-5.6 Sol |
| Older (GPT-5.6 era) | GPT-5.6 Luna | $0.20 / $1.20 | GPT-5.6 Luna |
| New (GPT-6 era) | GPT-6 Sol | $2 / $10 | This page |
| New (GPT-6 era) | GPT-6 Luna | $0.10 / $0.50 | This page |
OpenAI's wording is that it is “reducing API prices for Sol and Luna by 50% compared with their GPT-5.6 promotional pricing” — these are replacements for the earlier models of the same name, not new names living alongside them.
Pricing (USD per 1M tokens)
From the pricing table in OpenAI's announcement and the model pages in OpenAI's API docs.
| Model | Input | Output | Cached input | Change vs previous generation |
|---|---|---|---|---|
| GPT-6 Sol | $2.00 | $10.00 | $0.20 | $4/$20 → 50% cheaper |
| GPT-6 Luna | $0.10 | $0.50 | $0.01 | $0.20/$1.20 → 50% cheaper |
Cached input figures come from the API docs (Sol $0.20, Luna $0.01). The announcement text states “discounts of 90% on cached input-token reads”, which is consistent with 10% of $2.00 and 10% of $0.10.
Primary sources: OpenAI, “Introducing GPT-6 Sol and Luna” (pricing table) · GPT-6 Sol · GPT-6 Luna (verified 2026-09-25)
Figure: prices by generation
← scroll horizontally →
Specifications
Availability, verbatim
- “available in ChatGPT Work and Codex starting today for all Plus, Pro, Business, Enterprise, and Edu users”
- “Free and Go users can access GPT-6 Luna in the desktop app”
- “These models are not yet available in Chat”
- Rollout into ChatGPT happened progressively during the day; OpenAI advises checking back if you do not see them.
Sources: OpenAI API docs, GPT-6 Sol · GPT-6 Luna · announcement, Availability section
Six official benchmarks — with the operator named
Figures from OpenAI's announcement. OpenAI credits the source for each one, and more than half are third-party benchmarks. We add who runs each of them.
| Benchmark | Operated by | GPT-6 Sol | GPT-6 Luna |
|---|---|---|---|
| AutomationBench 1.0.6 | Zapier | (xhigh) 33.2% / $0.27 per task | (high) +5.4 pts vs previous generation, 58% lower cost per task |
| Agents' Last Exam V1 | agents-last-exam.org | (max) 56.4% — above Claude Opus 5's top score at 60% lower cost | — |
| FrontierCode 1.1 Main | Cognition | Large gain over GPT-5.6 Sol; matches Fable 5.1 (xhigh) at far lower cost | — |
| DeepSWE 1.1 | Datacurve | (max) 68.8% — within 1.1 pts of Fable 5's top 69.9% at ~80% lower cost | (max) 66.6% — comparable to Opus 5 / Fable 5 at medium effort |
| OSWorld 2.0 (offline, partial reward) | XLANG Lab | (xhigh) 60.5% vs Claude Opus 5 (medium) 60.3%, at ~80% lower cost | (max) beats GPT-5.6 Sol (medium) at one tenth of the cost |
| Factuality (internal eval) | OpenAI (internal) | About half the mistakes of the previous generation | At higher effort, matches GPT-5.6 Sol at roughly 1/100 of the cost |
Conditions behind each benchmark
| Benchmark | Conditions and notes |
|---|---|
| AutomationBench 1.0.6 | “AI agents are tested on end-to-end workflows using 47 tools across sales, marketing, operations, support, finance, and HR.” Comparison table: Sol (xhigh) 33.2% / $0.27, Astra (low) 30.3% / 3.9x, Claude Opus 5 (max) 26.9% / 11.1x, Fable 5.1 with Opus 5 fallback (max) 31.4% / over 8.9x. OpenAI's own caveat: “The datapoint for Claude Fable 5.1 understates its actual cost, as it omits the cost of the Opus 5 fallbacks, which occurred on ~40% of tasks.” |
| Agents' Last Exam V1 | “long-horizon, economically valuable tasks spanning 55 sub-industries” — covering most professional work performed on enterprise computers. |
| FrontierCode 1.1 Main | “graded not only on correctness but also ‘mergeability’: e.g., test quality, scope discipline, code style, and adherence to codebase standards”. |
| DeepSWE 1.1 | “AI agents solve original, long-horizon software engineering tasks” — original, not adapted from existing commits. |
| OSWorld 2.0 | “long-horizon computer-use workflows… We report the partial reward on the offline set from the v2026.08.08 release.” |
| Factuality | Internal evaluation built from “de-identified ChatGPT conversations where users had flagged a factual error”. OpenAI notes: “These error-inducing conversations are not representative of typical usage” and “Scores are not controlled for length.” |
Source: OpenAI, “Introducing GPT-6 Sol and Luna” (per-benchmark sections and caveats, verified 2026-09-25)
The other headline: better prompt caching
Alongside the price cut, OpenAI describes caching improvements for GPT-6: 90% off cached input reads, plus better default cache hit rates.
| Feature | From the announcement |
|---|---|
| Monitoring | A Prompt Caching Dashboard shows how much of your input was cached over time. |
| Diagnostics | A diagnostics tool explains missed cache opportunities and points at fixes. |
| Change settings without breaking the cache | Shifting reasoning effort or enabling/disabling tools keeps earlier context available for cache reuse. |
| Explicit breakpoints | Developers can choose where a cached prefix ends. |
| Real-world impact (GitHub) | GitHub reports that over the past few months it “reduced the share of prompt tokens requiring fresh processing by more than 50%” across billions of requests. |
Changing effort without invalidating the cache fits agent loops that dial reasoning depth up and down. Background on how caching works is in our glossary entry on prompt caching.
Source: OpenAI, “Introducing GPT-6 Sol and Luna” (Improving caching section)
Where these sit in the GPT-6 family
| Model | Price (in/out, 1M) | Role | On this site |
|---|---|---|---|
| GPT-6 Astra | $10 / $50 | OpenAI says it “continues to be our best model across the board” and to “choose it when you want the best results and an uncompromising experience”. Top tier. | Dedicated page |
| GPT-6 Sol | $2 / $10 | Hard business tasks with higher usage limits at lower cost. | This page |
| GPT-6 Luna | $0.10 / $0.50 | The cheapest practical tier; Free and Go users get it in the desktop app. | This page |
| GPT-5.6 Sol | $4 / $20 | Previous generation, replaced by GPT-6 Sol. Legacy. | Dedicated page (legacy note added) |
| GPT-5.6 Luna | $0.20 / $1.20 | Previous generation, replaced by GPT-6 Luna. Legacy. | Dedicated page (legacy note added) |
Astra's pricing follows our pricing comparison page (cache read $1.00). Astra was not part of this price cut and remains the top of the GPT-6 line.
Alignment and writing style
- Better than the GPT-5.6 versions: OpenAI says “both Sol and Luna show improvements over their GPT-5.6 counterparts, including lower rates of misleading claims about their coding work” — while noting that “the evaluations below deliberately test challenging situations and do not measure failure rates in typical use”.
- System card: published at deploymentsafety.openai.com/gpt-6-astra — the Astra card also carries the safety evaluations for this generation.
- Style inherited from Astra: “more clarity, less jargon, fewer odd turns of phrase, fewer low-value details, and slightly shorter answers overall without losing substance”, with the biggest gains in technical and coding conversations.
Source: OpenAI, “Introducing GPT-6 Sol and Luna” (Continuing to improve alignment, Collaboration style)
How to choose
- Cost-first bulk processing: GPT-6 Luna at $0.10 input / $0.50 output is near the front of the low-price band — see how it compares on our pricing comparison.
- Hard business tasks: GPT-6 Sol. OpenAI claims it beat Claude Opus 5 on AutomationBench at roughly one eleventh of the cost per task.
- Luna is not weak: on DeepSWE 1.1, Luna (max) scores 66.6%, comparable to Claude Opus 5 and Fable 5 at medium effort. It responds to more test-time effort.
- Watch for name confusion: check whether the URL is
/en/gpt-5-6-sol/or/en/gpt-5-6-luna/(legacy) versus/en/gpt-6-sol-luna/(current). Legacy pages keep the prices that were correct when published. - Peak accuracy: GPT-6 Astra remains the top model — see our GPT-6 Astra page.
Work out which of Sol and Luna is cheaper for you
Enter your input and output token counts to compare costs across every model we track.
🧮 Open the token calculator →📊 Full pricing comparison · 🤖 GPT-6 Astra · 🆚 Claude Opus 5.5 · 🗓️ Release timeline
⚠️ Disclaimer
- Pricing and specifications were verified on OpenAI's announcement and API documentation on 2026-09-25 and may change without notice. Always confirm on the official pages before you commit.
- Benchmark figures are OpenAI-reported, and competitor figures are quoted from each vendor's public reports. Some Claude comparisons are for Fable 5 rather than Fable 5.1. Conditions may not be identical and results are not guaranteed.
- This site is for information only and is not a recommendation or agent for any provider.