Two practical models designed to relieve the heavy pricing pressure of the computer-controlling flagship Astra have arrived. Here is an easy-to-read breakdown of the everyday all-rounder Sol, the ultra-cheap Luna, side-by-side pricing against competitors, and which one fits your daily workflow.
3-Line Summary
- On September 22, 2026, OpenAI officially announced GPT-6 Sol for general professional tasks and the lightweight GPT-6 Luna for high-volume automation, completing the GPT-6 lineup.
- Compared to the previous generation (GPT-5.6 Sol), API pricing dropped by nearly 50% ($2.00 per 1M input tokens for Sol), while Luna comes in at an astonishing 1/20th of Sol's price ($0.10 input).
- Teams can now cleanly split workloads: assign mouse-and-keyboard OS automation to Astra, complex document synthesis and coding to Sol, and hundreds of routine categorizations or summaries to Luna.
On September 3, OpenAI turned heads worldwide with GPT-6 Astra, a flagship model capable of watching your computer screen and operating software with a mouse and keyboard. Yet behind the technological wizardry lay a stubborn hurdle: cost.
Because Astra continuously captures and analyzes screen frames, it was tagged with a hefty price: $10.00 for 1M input tokens and $50.00 for output. Using it for drafting everyday emails or summarizing project memos was clearly overkill. Most knowledge work doesn't require an AI to physically grab your mouse; it needs sharp reasoning, clear writing, and reliable data interpretation.
Just two weeks later, on September 22, 2026, OpenAI launched GPT-6 Sol and GPT-6 Luna to address this practical reality.
Why are these models being hailed as a true price breakthrough, how are their roles divided, and which one should you reach for in your daily work? Let's break it down calmly.
Sol vs. Luna: How Do Their Roles Differ?
OpenAI clearly segmented these two models by workload complexity and budget:
[The GPT-6 Family Workload Split]
1. GPT-6 Astra ($10 / $50) ──> Visual screen understanding & OS mouse/keyboard actions
2. GPT-6 Sol ($2 / $10) ──> Proposals, data synthesis, software engineering, deep logic
3. GPT-6 Luna ($0.10 / $0.50)──> High-volume sorting, quick summaries, fast customer chat
1. GPT-6 Sol: The Everyday Intellectual Powerhouse
Sol is the dedicated "all-rounder" for professional knowledge work.
- It handles multi-page proposals, spreadsheet formulas, full-stack software development, and contract reviews—tasks where contextual nuance and rigorous logic matter most.
- Inheriting the recurrent depth (looped transformer) architecture proven in Astra, Sol dramatically reduces hallucinations and maintains coherent reasoning throughout long analytical tasks.
2. GPT-6 Luna: The Ultra-Low-Cost Speedster
Luna is built for pure throughput and rock-bottom operational costs.
- It shines in sorting hundreds of customer inquiries, extracting structured JSON from text, proofreading drafts, and handling straightforward chatbot FAQs—tasks where speed and cost per call outweigh deep philosophical thought.
- Billed at just 1/20th the cost of Sol, you can process thousands of records daily without sweating API bills.
Real-World Pricing Comparison
The headline feature of this release is unapologetic price cutting.
Here is how standard API rates (per 1M tokens) stack up across generations and rival labs:
| Model | Input Price (1M Tokens) | Output Price (1M Tokens) | Comparison vs. Flagships | Primary Sweet Spot |
|---|---|---|---|---|
| GPT-6 Astra | $10.00 | $50.00 | Flagship baseline | Screen recognition & OS navigation |
| GPT-6 Sol [New] | $2.00 | $10.00 | 80% cheaper than Astra (50% cut vs. 5.6) | Reasoning, strategic reports, coding |
| GPT-6 Luna [New] | $0.10 | $0.50 | 95% cheaper than Sol (Ultra-cheap) | High-volume sorting, summaries, chat |
| GPT-5.6 Sol (Predecessor) | $2.50 | $10.00 | Previous baseline | Everyday conversational tasks |
| Claude Opus 5.5 (Anthropic) | $9.00 | $45.00 | Heavyweight reasoning flagship | Deep architecture & complex research |
| Claude 3.7 Sonnet (Anthropic) | $3.00 | $15.00 | ~1.5x price of Sol | Balanced coding & agentic tasks |
| Gemini 3.8 Flash (Google) | $0.15 | $0.60 | Comparable lightweight tier | Rapid multimodal lookup & streaming |
Estimated Cost Across Real Scenarios
Token numbers can feel abstract until you apply them to tangible office tasks.
Scenario A: Auto-categorizing 1,000 incoming customer inquiries
(Average 300 input tokens per inquiry, 20 output tokens for classification / 300k input, 20k output tokens total)
- Using GPT-6 Astra: ~$4.00 — An obvious waste of money on simple text sorting.
- Using GPT-6 Sol: ~$0.80 — Decent, but still leaves room for efficiency.
- Using GPT-6 Luna: ~$0.04 (approx. 55 KRW) — Cheaper than a paper cup, processing 1,000 inquiries in seconds.
Scenario B: In-depth review and critique of a 50-page business proposal
(Full text + detailed prompt: ~40k input tokens, structured advisory response: 3k output tokens)
- Using GPT-6 Sol: ~$0.11 (approx. 150 KRW)
- For about 10 cents, you receive an executive-level breakdown of risks and strategic opportunities, delivering outstanding value for professional teams.
💡 Money-Saving Tip (Prompt Caching)
OpenAI introduced up to a 90% discount on cached input tokens with this release. If your application sends repetitive company guidelines, system instructions, or long reference documents, your real-world bill will drop far below the numbers listed above.
Which Model Should You Use?
With OpenAI's lineup now neatly tiered, matching the tool to your exact task delivers the best mix of speed and cost savings:
- Pick GPT-6 Astra: When the job requires visually scanning monitor screens, switching between multiple software windows, and automating physical mouse clicks.
- Pick GPT-6 Sol: When you need deep reasoning, flawless technical writing, financial modeling, or software engineering.
- Pick GPT-6 Luna: When dealing with bulk classification, extracting keywords from thousands of reviews, or building snappy customer-facing chatbots.
Frequently Asked Questions
Can ChatGPT Plus subscribers use Sol and Luna right now?
Yes. ChatGPT Plus, Pro, and Business users can pick GPT-6 Sol as their default daily model from the top selector, or switch to Luna whenever rapid responses are needed.
Is there a noticeable leap over the previous GPT-5.6?
The difference is most palpable in logical consistency across long conversations. While previous models sometimes lost track of earlier constraints when analyzing hefty documents, Sol's looped transformer architecture keeps arguments coherent until the final line.
Is Luna too cheap to write natural prose?
For standard summaries, translations, and everyday conversational English and Korean, Luna is surprisingly fluid. However, for nuanced legal interpretations or complex multi-step math proofs, Sol remains the safer choice.
Can free users access these models?
OpenAI has begun rolling out GPT-6 Luna to free and Go tier users within the desktop application. Sol access for free users is offered on a limited daily trial quota.
Can we combine Sol and Luna in a single API pipeline?
This is exactly how many production apps are built. Send every incoming user message to Luna ($0.10) for initial intent routing. If it's a simple greeting or FAQ, Luna answers directly. If it requires multi-step analysis or code execution, escalate the query to Sol ($2.00). This hybrid design routinely slashes API expenses by over 70%.