OpenAI packed a year's worth of news into one day last week.
The company released GPT-5.6 on July 9th, a new AI generation that ships as three separate models named Sol, Terra, and Luna, each tuned for a different kind of work.
Plus, and it surrounded them with new places to use them: ChatGPT Work, an agent that carries whole projects from goal to finished deliverable; an expanded role for Codex, its technical agent; and GPT-Live, a rebuilt voice mode that converses like a phone call. Together the releases change how a dealership should think about AI.
Here is the breakdown, as well as a hands-on test your team can run this week to get acquainted with the updates:
GPT Gets a New Family
GPT-5.6 is the latest generation of OpenAI's flagship AI, the engine behind ChatGPT. It rolled out to the public on July 9, 2026 after a brief invite-only preview that OpenAI coordinated with the U.S. government (a first, tied to new federal rules around powerful AI models).
☀️ Sol: The Flagship
What it is: OpenAI's most capable model ever, built for the hardest, longest, multi-step problems — deep analysis, complex coding, research, and long-running projects.
When to use it:
Analyzing a month or quarter of sales, service, or F&I data and asking "what's really going on here?"
Building something from scratch: a full BDC training program, a pay-plan comparison, a market pricing analysis
Any task where you'd say "this needs to be right, and I'll wait for it"
Cost (API pricing): $5 per million input tokens / $30 per million output — the premium option.
🌎 Terra: The Everyday Workhorse
What it is: The balanced middle tier. Performs about as well as last generation's flagship at roughly half the price. OpenAI positions it for high-volume business work: customer support, document analysis, internal tools.
When to use it:
Rewriting used-vehicle descriptions and merchandising copy
Summarizing customer reviews, CSI surveys, or recorded-call notes
Drafting objection-handling scripts, email templates, and service-drive word tracks
Basically, if you're not sure which model to use, start here.
Cost: $2.50 / $15 per million tokens — half of Sol.
🌖 Luna: The Speed Runner
What it is: The fastest, cheapest member of the family. Built for quick, high-volume, routine tasks where speed matters more than deep thinking.
When to use it:
First-response texts and emails to internet leads (speed-to-lead is everything)
Quick summaries: "Boil this trade appraisal disclosure down to three bullets"
Labeling, sorting, and extracting like pulling names, VINs, and appointment times out of messy notes
Routine drafts you'll polish yourself anyway
Cost: $1 / $6 per million tokens — a fifth the price of Sol.
One access note: In the regular ChatGPT app, paid users get Sol through the reasoning settings (GPT-5.5 Instant still handles quick everyday replies). All three models are selectable in ChatGPT Work, Codex, and the developer API.
How It's Different From Previous GPTs
Here's what actually changed from GPT-5.5 and earlier versions:
It does more work per dollar. Terra performs roughly on par with the previous flagship (GPT-5.5) at about half the cost. For any store paying for AI tools — or any vendor whose products run on OpenAI under the hood (many dealership chat, BDC, and marketing platforms do) — that math eventually flows downstream to you in the form of cheaper or more capable tools.
It's built for real work documents. OpenAI is pushing GPT-5.6 hard as a "get work done" model, not just a chatbot. It's notably better at producing editable presentations, spreadsheets, documents, and polished layouts, and at pulling information out of reference files you upload. Translation: hand it your month-end numbers or a stack of CSI survey comments and get back something you can actually present at the Saturday meeting.
It thinks harder when you ask it to. GPT-5.6 Sol adds a new "max" reasoning setting for genuinely hard problems, plus an "ultra" mode (on higher-tier plans) that splits big projects across four AI agents working in parallel. You probably won't touch ultra mode for a follow-up text—but for something like analyzing a full quarter of deal data, it's a different class of tool.
It's faster and less wordy. OpenAI says the new models complete complex tasks in roughly half the time using far fewer words than before. If you've ever asked ChatGPT a simple question and gotten a term paper, you'll appreciate this.
You pick the model, not just the answer. This is the philosophical shift. Older GPT releases were one-size-fits-all. GPT-5.6 asks you to match the model to the job.
Source: OpenAI
Work, Codex, and Live
The models arrived alongside three new products that put them to use:
What it is: An agent built into ChatGPT that accepts a goal, breaks it into steps, and keeps working for hours until the job is finished. It gathers context from connected apps like Gmail, Google Drive, Slack, and calendars, then produces finished documents, spreadsheets, presentations, and reports.
Why the design matters: It runs on a cloud machine that stays on, so a manager can assign a project from a phone between customers and review the output later, and Scheduled Tasks let it repeat work automatically.
Put it to work: Hand it a month-end assignment like "pull the numbers from this DMS export, compare them to last quarter, and build the Saturday-meeting deck," or schedule a Monday task that summarizes every new Google review and drafts a response to each. Early corporate users report weeks-long analysis cycles compressing into hours.
Availability: Rolling out across paid plans, starting with Pro and Enterprise and reaching Plus and Business within days. It is also the surface where all three GPT-5.6 models sit side by side in the picker (which makes it ideal for the test below).
What it is: OpenAI's coding agent, now merged into a new ChatGPT desktop app. More than five million people use it weekly, over a million of them for work outside software, which persuaded OpenAI to build its technology directly into ChatGPT Work.
Why your store should care: Your website, chat, and CRM vendors use agents like Codex to ship improvements faster, so its gains reach you through better software, and a store with modest technical skill can use it to build small internal tools, like a script that flags inventory photos missing from listings.
What it is: A new voice model that listens and speaks at the same time, the way a phone call works, replacing the old turn-based voice mode. It holds natural conversations, handles interruptions, supports live translation, and hands harder questions to a frontier model in the background while the conversation keeps flowing.
Availability: GPT-Live-1 becomes the default voice for paid ChatGPT users, with a mini version serving the free tier.
Put it to work: Talk through a deal structure hands-free while walking the lot, rehearse objection handling out loud on the drive in, or translate live for a customer whose first language differs from yours.
One Prompt, Three Outputs
As LeVar Burton would say, “Don't take my word for it.”
Below is a test you can try yourself.
Run the identical prompt through Luna, then Terra, then Sol inside ChatGPT Work to evaluate how differently they rationalize and process requests.
You are the sales manager at a mid-size dealership. A customer named Dana emailed tonight:
"I test drove the [YEAR, MAKE, MODEL] on Tuesday and I liked it, but I'm $4,200 upside down on my trade, CarMax offered me $2,000 more than you did, and another store 40 minutes away quoted me a lower out-the-door price. I'd rather buy from you since you're close to my house, but you need to earn it. What can you do?"
Do three things:
Analyze the situation: identify what Dana is really telling us, rank the three objections by how much they actually threaten the deal, and explain your reasoning.
Lay out a response strategy: what we should concede, what we should hold firm on, and what we should reframe, with the goal of an appointment tomorrow rather than a bidding war by email.
Write the reply email itself: in under 150 words, warm and confident, ending with a specific appointment ask.
What to Watch For
Each model will complete the assignment, and the differences will show in how well each part holds up.
On the analysis, look for whether the model reads between the lines, because Dana's message contains a buying signal ("I'd rather buy from you") that a sharp manager weighs heavier than any objection.
On the strategy, look for judgment: a strong answer distinguishes the CarMax number from the competing store's quote and explains why they call for different moves.
On the email, look for something you would send tonight with your name on it.
Expect Luna to return a fast, usable email attached to a surface-level strategy, Terra to deliver a solid version of all three parts, and Sol, especially at higher reasoning effort, to separate on the analysis and strategy with a sharper read of Dana's leverage.
The Scorecard
Print the three outputs with the model names removed and put three questions to your team:
Which email would you send tonight?
Which strategy would you run tomorrow?
Which analysis told you something about Dana you had missed?
When the votes split across outputs, you have watched the tier system work: each model earned its answer at a different price, and the exercise shows which level of thinking each of your store's daily tasks requires.
Match this month's workload to those levels and you have a routing plan, built from evidence, in one afternoon.
The Bottom Line
GPT-5.6 turns AI into a menu, and the stores that win with it will be the ones that match the right model to the right job, the same way they already match the right person to the right customer.
Luna covers speed, Terra covers the daily grind, and Sol covers the heavy lifting. Run the one-prompt test and see where each one fits in your store.
