Hello There!
See your building as a living 3D twin you own.
Hello There!
See your building as a living 3D twin you own.
Scope3D Desktop has no AI of its own. It drives the account you already hold,
and you are billed by that provider at their published rates.
A fixed list, not a fetched one. The picker asks which provider first, because that is the account being billed, and every model carries the same two numbers the app charges you against.
| Provider | Model | When to pick this | Rate, per million tokens |
|---|---|---|---|
| Anthropic | Opus 5 · default | Long traces and multi-step work. What the prompts are tuned against. | $5 in · $25 out |
| Anthropic | Sonnet 5 | Near-Opus quality, faster and cheaper. Good for edits and questions. | $3 in · $15 out |
| Anthropic | Fable 5 | The most capable model here, for a drawing the others misread. Slowest and priciest. | $10 in · $50 out |
| Anthropic | Opus 4.8 | The previous Opus. Useful for comparing a run against a known baseline. | $5 in · $25 out |
| Anthropic | Haiku 4.5 | Fastest and cheapest. Fine for lookups, not for tracing. | $1 in · $5 out |
| OpenAI | GPT-5.6 Sol | The frontier tier, 1M context. Where a run goes if Anthropic is down. | $5 in · $30 out |
| OpenAI | GPT-5.6 Terra | Most of Sol's capability at under half the price. Edits and questions. | $2 in · $12 out |
| OpenAI | GPT-5.6 Luna | Cheapest of the three. Lookups, not tracing. | $0.20 in · $1.2 out |
| Gemini | Gemini 3.1 Pro (custom tools) · preview | Pro reasoning on the endpoint tuned to prefer custom tools, which is all this app calls. | $2 in · $12 out |
| Gemini | Gemini 3.1 Pro · preview | The same model on the general endpoint. Strongest of the three at reading a drawing. | $2 in · $12 out |
| Gemini | Gemini 3.6 Flash | Fast, cheap and — unlike the Pro pair — not a preview. 1M context. | $1.5 in · $7.5 out |
| Moonshot | Kimi K3 | Moonshot's frontier model. Always reasons, and the effort control sets how hard. | $3 in · $15 out |
| Moonshot | Kimi K2.7 Code | Tuned for code, which is most of what this app asks a model to write. | $0.95 in · $4 out |
| Moonshot | Kimi K2.7 Code (fast) | The same model tuned for speed. Edits and questions, not tracing. | $1.9 in · $8 out |
Rates as of 2026-08-04, taken from each provider's own published table and printed beside the model inside the app. Anything added later without that check is flagged in the app as an estimate rather than quietly shown as fact. Gemini prices in two bands either side of 200K tokens and the app models the upper band separately — a floor-plan trace passes 200K within a few steps and then stays there, so one flat rate would under-report most of a long trace.
The Sonnet 5 line is the standard rate, not the $2 / $10 introductory one that runs to 2026-08-31: a rate that expires in weeks would leave the table under-reporting spend from September, and over-stating for a few weeks is the safer direction to be wrong. Cached tokens are not billed flat either — Anthropic reads a cached prefix at a tenth of the input rate and writes one at 1.25×, and the app applies both multipliers rather than charging you input price for a prefix it re-sent.
This is what Scope3D spent, not your account balance. None of the four providers will tell an ordinary API key what the balance is, so the app measures its own consumption exactly and links to their console for the authoritative figure rather than inventing one.
The app used to decide for itself. A run started on Opus 5 could fail over to Gemini part-way through and finish as a model nobody had chosen — and every figure taken from it afterwards named the wrong one.
A run that dies is obvious and repeatable. A run that quietly finishes as somebody else looks exactly like success. So the app keeps its opinion and loses its authority: it classifies the failure, says whether trying again can plausibly work, and waits to be told what to do.
Automatic failover was removed on purpose
Nothing moves providers without you. It also never switches on a malformed request or a refusal — one is our bug and the other is a decision, and changing model would only hide them behind a different error message.
The run pauses and names the failure
Which model could not answer, the provider's own words kept verbatim, which attempt at this turn it was, and where the run had got to — "stage 2 of 6, step 11".
Retry, Stop, or switch to any of the other thirteen
Retry re-asks the same model with the same messages, so a capacity blip costs only the wait. Models you hold no key for are listed and disabled rather than hidden — a shorter list would leave you believing you have fewer choices than you do.
Nothing is lost while it waits
A switch happens at the turn boundary, and the previous model's private reasoning is stripped there because each provider signs its own. Everything said out loud survives the move.
The short version: the bill is your AI provider's, and the only ceiling in the app is the one you type into the budget box.
Licence key
There isn't one. Nothing to enter, nothing to lose, nothing to reissue when you get a new laptop.
Trial
There isn't one either. Nothing expires and nothing unlocks later, so there is no countdown to plan a pilot around.
Seats
None. It is not a per-user product, so there is no user list to be added to or removed from.
Usage cap
The only spending limit in the application is the per-provider budget you set yourself, and you can change or clear it whenever you like.
What you pay
Your AI provider, directly, at their published rates. We never resell usage and never take a cut, which is also why the app has to measure spend rather than bill it.
Attribution
Revit access is provided by rvt-mcp, licensed under Apache-2.0 and redistributed unmodified. The canonical licence and attribution text ships beside the executable and is reachable at the foot of the Setup screen, because the obligation is to the person running the software rather than to a repository.
If yours is not here, ask us and we will answer it properly.
Yes. Pick from the chip in the header — provider first, then models — and the transcript carries across. The one thing that does not travel is the previous model's private reasoning, which each provider signs, so it is dropped at the turn boundary. Everything said out loud stays. A sensible pattern is to trace on the top tier and drop down to a cheap model to ask questions about what you built.
In Windows Credential Manager. A pasted key goes straight through to the application's Rust side and is signed into requests there — it is never held in component state and never rendered back into the window. Remove this key deletes it from Credential Manager; revoking it at the provider is a separate step, and their console is linked from the same place.
It stays listed and is labelled "no key". One key is enough to work; the other three are there so you can see what connecting them would give you. Setup shows each provider as its own row with a status line and the button that fixes it, and it re-checks the machine while you work, so a step ticks itself once the key lands.
No. The browser twin plans are for the hosted twin — the model an owner opens, explores and keeps for the life of the building. Scope3D Desktop is not on that ladder. The two products meet at the file, not at the invoice.
One key for any of the four providers is enough to start. We will run a floor with the Spend drawer open, so you see the real number rather than an estimate.