GPT Models for Contractors: A Plain-English Guide
A breakdown of GPT-5.6 for general contractors: what Sol, Terra, and Luna do, how effort levels work, and which model fits each precon task.
Michael Sullivan
Senior Growth Marketer
You boot up ChatGPT to knock out a quick email, and the picker has turned into a menu. First the model: GPT-5.6, now split into three tiers named Sol, Terra, and Luna. Then an effort setting with options like medium, max, and ultra. Somewhere a coworker told you Sol is the good one, and now you’re standing in the tool aisle holding three things that look almost identical, reading the fine print, wondering which one you actually need.
The good news: you already know how to make this call. You match your effort to the stakes of the decision every day on the job. You do not run a full risk review to pick a paint color, and you do not wing a $30M pursuit. Same idea here. Match the model to the stakes of the task, and the picker stops being intimidating.
What is GPT-5.6? GPT-5.6 is the current family of general-purpose AI models behind ChatGPT: they read and write plain language, summarize documents, and answer questions. The family comes in three tiers, Sol, Terra, and Luna, and you steer it with two dials. The model tier sets the base capability and the cost. The effort level sets how hard the model thinks before it answers. That is the whole control panel.
Key Takeaways
- GPT-5.6 gives you two dials: which tier you pick, Sol, Terra, or Luna, and how much effort you let it spend.
- Effort levels scale the same model up: medium is the default, max buys more thinking, and ultra puts several agents on the problem at once.
- Route by stakes: reach for Luna or Terra on everyday writing and lookups, and save Sol or Sol Pro for your hardest analysis.
- Contract clauses and code questions call for Sol at max effort plus human verification; the model is often confidently wrong there.
- Bid leveling, go/no-go, and cost benchmarking need your own data, so no tier or effort level fits; that is Kit’s job.
The GPT-5.6 Family, Broken Down
On July 9, 2026, after a limited preview, OpenAI released GPT-5.6 to everyone. The pitch is more intelligence from every token: better answers per dollar, with capability you can dial up when the work gets hard. The number, 5.6, is the generation. The three names are durable capability tiers, each free to improve on its own schedule.
Here is what each tier is built for:
| Tier | Built for | The short version |
|---|---|---|
| Sol | Your hardest work | Flagship capability across coding, analysis, and judgment. Turns messy notes into polished documents. A Sol Pro variant handles the most complex tasks. |
| Terra | Everyday professional work | The balanced daily driver, priced well below Sol, with performance in the neighborhood of the last generation. |
| Luna | High-volume, low-cost work | The fastest and cheapest tier, built to make a good answer cheap enough to use all day long. |
The prices tell the same story. Measured per million tokens of text in and out, the tiers ladder down in cost:
| Tier | Input (per 1M tokens) | Output (per 1M tokens) |
|---|---|---|
| Sol | $5 | $30 |
| Terra | $2.50 | $15 |
| Luna | $1 | $6 |
Most of your team pays for ChatGPT by subscription, not by the token, so treat these numbers as a signal rather than a bill. Terra runs about half the cost of Sol; Luna about a fifth. That gap is the whole reason you would not send every quick email to the flagship.
What you can actually pick depends on where you are working. In everyday ChatGPT, paid plans run on Sol and let you set the effort, and Pro and Enterprise add Sol Pro for the highest-quality answers. In ChatGPT Work and Codex the menu opens up: Free and Go plans get Terra, and paid plans choose freely among Sol, Terra, and Luna. The takeaway for a busy precon team: you almost always have a cheaper, faster tier sitting right there for the easy stuff.
Effort Levels: Capability on Demand
The second dial is the one people miss, and it is the more interesting half of the story. Effort is how much thinking budget you hand the model before it answers. Same model, more or less horsepower.
Three settings matter:
- Medium is the efficient default: fast, cheap, and fine for most work.
- Max hands the model more time to reason, explore alternatives, check its own work, and revise before it answers. Think of the difference between a quick eyeball estimate and a full quantity takeoff with a second set of eyes on the math. Same estimator, a lot more rigor.
- Ultra goes further still, putting several copies of the model on the problem at once and reconciling their answers, the way you might put three estimators on the same takeoff and settle on the number together. It costs more tokens and more time; you spend it on genuinely hard problems.
None of this changes which model you are talking to. It changes how much work that model does before it speaks. The tradeoff is always time: max and ultra think longer, so you feel the wait. Spend it where the answer is worth waiting for. A careful Sol at ultra has no business drafting a one-line email, and a quick Luna at medium has no business pricing your riskiest scope.
Match the Model to the Task
Most precon work a general chatbot can help with is low-stakes writing and lookups. A few pieces are high-stakes reasoning. And the highest-stakes work does not belong in a general model at all. This is only the picker; for the wider view of how general contractors are putting AI to work in 2026, we have a separate guide. Here is the map:
| Precon task | Reach for |
|---|---|
| Sub nudge and reminder emails, invitation to bid cover notes, BD outreach, meeting recaps, quick definitions, unit conversions | Luna or Terra, medium effort |
| Summarize a spec section, digest an addendum, pull dates and bonding requirements from bid docs, draft RFI questions, gut-check a $/SF number | Terra or Sol, medium to max, and you review it |
| Plain-English a contract clause, risk-assess unusual terms, building code questions, make the go/no-go case to leadership | Sol at max or ultra, or Sol Pro, and verify everything |
| Bid leveling, go/no-go on real numbers, cost benchmarking against your history | No tier or effort fits. This needs your data. |
Where Buildr fits: that bottom row is not a setting you can dial up. No tier and no effort level helps a model that has never seen your data. More on that below.
Picture the week. It is 11pm and you are reading a ninety-page spec section you did not write, hunting for the submission and insurance requirements buried in it. That is a Terra job at max effort: let the model pull the key points into plain English, then you confirm them against the source. The next morning you nudge four subs who have gone quiet on an invitation to bid. Pure Luna work; a fast, cheap draft you tweak and send.
Then Thursday night rolls around, bid day is tomorrow, and you are leveling five subcontractor quotes to figure out who carried the temp power and who left it out of their number. No general model can touch that, no matter which tier or effort you click. It has never seen these subs or these scopes. And when you paste a termination clause in and ask what your exposure is, treat the answer as a first pass only: run Sol at max effort, then read the actual contract yourself, because the model is often confident and wrong on legal and code questions.
The pattern holds across all of it: the model does the tedious first pass, and the estimator still owns the judgment and the number.
Your precon data lives
in 6 places.
One AI-powered platform for your CRM, estimates, workforce planning, and bids. No more tab-switching.
The Line a General Model Cannot Cross
Every general GPT model, at every tier and effort level, shares one hard limit: it starts blank every session. It knows the internet. It does not know your business. It has never seen your historical budgets, your win/loss record, your sub performance, or the fifteen years of pricing sitting in your files.
That is fine for words. It is a dealbreaker for numbers. Leveling bids, scoring a go/no-go against your real pipeline, benchmarking an estimate against what you actually paid last time: none of that is a language problem you can solve by thinking harder. It is a data problem, and the answer has to come from your data, not from a well-read stranger who has never worked for you.
That is the whole reason fit-for-purpose tools exist. We made the fuller version of this case in our post on ChatGPT in construction, and it runs through everything we write about AI in construction: the tool is only as sharp as the data behind it.
What Kit Does That the Best Model Can’t
That fit-for-purpose tool has a name. Kit is the agentic assistant built into the Buildr Platform, and the point is not that it out-thinks GPT-5.6 Sol. The point is that it works inside your business instead of starting from a blank page. None of the following is a setting you can unlock on a general model, at any tier or effort level:
- Read your actual invitation to bid and pull the scopes. Upload the document and Kit breaks out what each trade covers, the key dates, and the submission requirements. An hour of reading at 11pm becomes a two-minute review you sign off on.
- Level bids against your own scope list. Ask Kit to compare submissions side by side and it flags who included the fireproofing and who quietly left it out, mapped to the scopes you actually track.
- Answer questions about your pipeline. Win rate by market sector, high-probability pursuits sorted by award date, repeat clients who have gone quiet with no active lead: Kit reads your CRM, pulls the report, and charts it if you ask.
- Benchmark against what you actually paid. Kit checks a number against your historical budgets, so “average cost of lumber per square foot on industrial jobs” comes from your jobs, not the internet’s guess.
- Learn how your team works. Correct it once and it remembers. The first ITB might need cleanup; by the fifth it is dialed into your preferences. A public chatbot forgets you the moment you close the tab.
- Keep your data yours. Kit runs on your information inside your environment, so sub pricing and owner budgets never get pasted into a public tool you do not control.
Like any good teammate, Kit works best as a conversation, not a vending machine. Ask, look at what it gives you, then tell it what to fix: “focus on mechanical and electrical” or “that square footage looks off, check it again.” Treat it like a fast junior estimator who needs your direction, and it sharpens the more you steer it.
So keep using ChatGPT. It is genuinely good at the writing that eats your evenings. Just know where the line is: general AI is great for words, and your numbers, your documents, and your decisions belong to a tool built on your data. The model is a sharp tool. Your estimator is still the one who builds the bid.
FAQ
What is the difference between a GPT-5.6 model and an effort level?
The model tier, Sol, Terra, or Luna, sets the base capability and cost: Sol is the flagship, Terra is the balanced middle, and Luna is the fast, cheap option. The effort level, from medium up to max and ultra, sets how hard that model thinks before it answers. Pick the tier for the job, then dial the effort to the stakes.
What do max and ultra effort do?
Max gives the model more time to reason, check its own work, and revise before answering. Ultra goes further and runs several copies of the model in parallel, then combines their answers for a stronger result. Both trade time and cost for quality, so save them for genuinely hard problems.
Which GPT-5.6 model should I pick, Sol, Terra, or Luna?
Use Luna for high-volume, low-stakes writing and lookups, Terra for everyday professional work like summaries and drafts, and Sol for your hardest analysis. Pro and Enterprise plans can step up to Sol Pro when the answer really has to be right.
Can ChatGPT level subcontractor bids or run a go/no-go decision?
No. Bid leveling, go/no-go scoring, and cost benchmarking all depend on your subs, your numbers, and your pipeline history. A general model has never seen any of it and starts blank every session. That work needs AI built on your connected data.
What model and effort should I use for a contract clause or code question?
Use Sol at max or ultra effort, or Sol Pro if you have it, and then verify the answer against the actual contract or code. These are high-exposure questions where a general model is often confident and wrong, so it does the first pass and a person owns the final call.
What can Buildr’s Kit do that ChatGPT cannot?
Kit works inside your Buildr data, so it can read your uploaded invitation to bid and pull the scopes, level bids against your own scope list, answer questions about your pipeline and win rate, and benchmark costs against your historical budgets. A general model like ChatGPT has none of that context and starts blank every session.