What are GPT-6 Sol and Luna? GPT-6 Sol and Luna are OpenAI's new mid-tier and low-cost models in the GPT-6 family, released on September 22, 2026 alongside a 50% API price cut versus their GPT-5.6 predecessors. For CRE investors, GPT-6 Sol and Luna matter less as a benchmark story and more as a cost story: document work that used to require budgeting, like abstracting every lease in a portfolio or screening every offering memorandum that hits your inbox, now costs cents per document. For the broader tool landscape, see our guide to the best AI tools for commercial real estate investors.
Key Takeaways
- GPT-6 Sol now costs $2 per million input tokens and $10 per million output tokens, half the GPT-5.6 Sol promotional price of $4 and $20.
- GPT-6 Luna costs $0.10 per million input tokens and $0.50 per million output tokens, putting a full 60-page lease abstract at roughly half a cent.
- Cached input tokens are discounted 90%, which makes repeated questions against the same lease, OM, or rent roll dramatically cheaper for CRE teams.
- OpenAI reports Sol beats Claude Opus 5 on a business-workflow benchmark at about 9% of the cost per task, though these are vendor-reported results.
- For CRE investors, the practical shift is from sampling documents to processing all of them, since per-document AI costs are now trivial.
GPT-6 Sol and Luna Explained
OpenAI launched GPT-6 Astra earlier in September 2026 as its flagship, which we covered in what GPT-6 Astra's computer use means for CRE investors. GPT-6 Sol and Luna are the next two tiers down. According to OpenAI, both were trained with methods similar to Astra's, carrying its gains in professional work, factuality, coding, and computer use into faster, cheaper models. Astra remains OpenAI's recommendation when you want the best results regardless of cost.
The three tiers map cleanly onto how a CRE firm actually spends AI budget:
- GPT-6 Astra: the hardest, highest-stakes work, such as a complex joint venture waterfall review or a contested lease interpretation.
- GPT-6 Sol: the daily workhorse for underwriting support, OM review, investor memo drafting, and agentic workflows across apps.
- GPT-6 Luna: high-volume extraction and classification, such as pulling rent, term, and escalation fields from thousands of documents.
If you tracked the previous generation, our breakdown of GPT-5.6 Sol, Terra, and Luna costs for CRE investors shows how fast this pricing curve has moved since July.
The New Pricing, and What It Buys a CRE Firm
Here is the published GPT-6 API pricing, per 1 million tokens:
- GPT-6 Sol: $2 input, $10 output (was $4 and $20 for GPT-5.6 Sol)
- GPT-6 Luna: $0.10 input, $0.50 output (was $0.20 and $1.20 for GPT-5.6 Luna)
Abstract pricing means little until you run it against real documents. Using a rough estimate of 40,000 input tokens for a 60-page commercial lease and 3,000 output tokens for a structured abstract:
- One lease on GPT-6 Sol: about $0.08 of input plus $0.03 of output, roughly $0.11 per lease.
- One lease on GPT-6 Luna: about $0.004 plus $0.0015, roughly half a cent per lease.
- A 1,000-lease portfolio: roughly $110 on Sol or about $5.50 on Luna, versus roughly $220 on GPT-5.6 Sol at its prior price.
- Screening 300 offering memorandums a month on Sol: at roughly 35,000 input and 2,000 output tokens each, about $0.09 per OM, or around $27 a month.
Your actual token counts will vary with document length, scan quality, and prompt design, but the order of magnitude is the point. At these prices, AI processing is no longer the expensive part of a lease audit or an acquisitions screen. Human review time is.
Why Caching Matters More Than the Headline Price
The less obvious change is prompt caching. OpenAI says GPT-6 delivers higher cache hit rates by default, with cached input reads discounted 90%. It also now lets developers change reasoning effort and toggle tools without breaking the cache.
CRE work is unusually cache-friendly because analysts ask many questions of the same document. Consider an asset manager asking 20 follow-up questions about one 40,000-token lease on Sol. Without caching, the input cost is about $1.60. With the first read at full price and 19 cached reads at $0.20 per million tokens, the input cost drops to roughly $0.23. The same logic applies to an agent that repeatedly checks a rent roll, a T-12, and a loan agreement while it works.
That is what makes always-on monitoring realistic. An agent can re-read a property's operating statement every month and flag drift. For example, if NOI falls from $1.2 million to $1.05 million against $900,000 of annual debt service, DSCR moves from 1.33x to about 1.17x, which may trip a covenant before anyone opens the spreadsheet. At a $20 million valuation, the same NOI drop takes the implied cap rate from 6.0% to 5.25%, a gap that should prompt a hard look at the carrying value.
What the Benchmarks Say, and How to Read Them
OpenAI published several comparisons. All of the figures below are vendor-reported, and competitor scores were taken from public reports:
- AutomationBench (business workflows across 47 tools): GPT-6 Sol at xhigh effort scored 33.2% at $0.27 per task, versus 26.9% for Claude Opus 5 at max effort, which cost 11.1 times as much per task.
- Agents' Last Exam (long-horizon professional work across 55 sub-industries): Sol scored 56.4% at max effort, above Opus 5's best score at 60% lower cost per task.
- OSWorld 2.0 (computer use): Sol at xhigh scored 60.5% versus 60.3% for Opus 5 at medium, at about 80% lower cost per task.
- Factuality: on an internal evaluation built from conversations where users flagged errors, Sol made about half as many mistakes as its predecessor.
Treat these as a reason to test, not a reason to switch. Anthropic's Claude, Google's Gemini, and OpenAI's models each have strengths, and the right answer for a lease abstraction pipeline depends on how each model handles your documents, your scans, and your edge cases. Our lease abstraction tools comparison lays out how to run that bake-off fairly.
Real-World CRE Applications
The shift from "sample the documents" to "process all of them" opens up workflows that were technically possible before but hard to justify on cost:
- Full-portfolio lease audits: run every lease through Luna for field extraction, then route only the exceptions, like unusual co-tenancy clauses or CAM caps, to Sol or a human reviewer.
- Every-OM screening: score every inbound offering memorandum against your buy box instead of skimming the ones a broker flags. See our walkthrough on AI offering memorandum review for a workflow you can adapt to any model.
- Monthly covenant watch: cache each loan agreement once and let an agent check DSCR and debt yield tests against each new operating statement.
- Diligence room triage: classify and summarize hundreds of data room files on day one of a purchase contract, not day ten.
The constraint is rarely the model anymore. It is the workflow design, the verification step, and getting your team to trust the output. That is where most firms stall: JLL's 2025 Global Real Estate Technology Survey found that 92% of corporate real estate teams have started or planned AI pilots, but only 5% report achieving most of their program goals. CRE investors looking for hands-on AI implementation support can reach out to Avi Hacker, J.D. at The AI Consulting Network.
Availability and How to Start
According to OpenAI's announcement, GPT-6 Sol and Luna are available in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu users, rolling out gradually. Free and Go users can access Luna in the desktop app. The models are not yet available in standard Chat. In the API they are named gpt-6-sol and gpt-6-luna.
A sensible first step: pick 20 leases you have already abstracted by hand, run them through Luna and Sol, and compare field-level accuracy against your manual abstracts before you scale anything. For personalized guidance on implementing these strategies, connect with The AI Consulting Network.
Frequently Asked Questions
Q: What is the difference between GPT-6 Sol and GPT-6 Luna?
A: GPT-6 Sol is OpenAI's mid-tier model for complex professional work at $2 input and $10 output per million tokens. GPT-6 Luna is the low-cost tier at $0.10 and $0.50, best suited to high-volume extraction and classification tasks like pulling fields from leases.
Q: How much does it cost to abstract a lease with GPT-6 Luna?
A: Assuming roughly 40,000 input tokens for a 60-page lease and 3,000 output tokens for the abstract, GPT-6 Luna costs about half a cent per lease. A 1,000-lease portfolio would cost around $5.50 in API fees, before any human review.
Q: Should CRE firms switch from Claude or Gemini to GPT-6 Sol?
A: Not on benchmarks alone. OpenAI's comparisons are vendor-reported, and model performance on your own leases, OMs, and rent rolls is what matters. Run a side-by-side test on documents you have already reviewed by hand, then decide per workflow.
Q: Is GPT-6 Sol accurate enough for underwriting work?
A: OpenAI reports Sol makes about half as many factual errors as its predecessor, but no model should produce final underwriting numbers without verification. Use it to extract, summarize, and flag, then have an analyst confirm NOI, cap rate, and DSCR inputs before they reach an investment committee.