Using Rapid.Coder
Rapid.Coder is the AI that writes your dashboards. Here is how to pick the right model and keep the bill sensible.
What Rapid.Coder is
Every prompt you type goes to an AI model that reads your data, writes the queries, and assembles the page. Rapid.Coder is the shortlist of models available to your box, along with the benchmark scores and live pricing for each one. It exists so you can pick a model in ten seconds rather than researching the field.
A model earns its place here on human votes, not test scores. Both
WebDev Arena and
Design Arena show people two results side by side
without saying which model made which, and ask which is better. That is a closer match to how you will judge a
dashboard than any automated benchmark. We then use the three
Artificial Analysis scores and
OpenRouter pricing to choose between models people
rate similarly. The full catalog lives on the Rapid.Coder catalog page,
and the machine-readable version is at rapid.coder/model.json.
You can build all day without touching this page. Your box ships with a sensible default already selected. Come here when a build feels slower than it should, when a hard problem needs more horsepower, or when someone asks what the AI costs.
Code or design: two rankings
Building an application that works and designing one that looks right are different jobs, and the models are not equally good at both. So every model carries two headline scores.
| Score | Answers | Measured by |
|---|---|---|
| Code | Does the thing actually work? | Automated test suites. Whether code runs is objectively checkable, so a machine is the right judge. |
| Design | Does it look like something you would show a client? | Blind human voting. People are shown two results without being told which model made either, and pick the better one. The score is the percentage of those matchups the model won. |
Sorted by code score. Notice the order does not hold: Kimi K3 out-designs Claude Opus 5 despite scoring lower on code, and costs less. DeepSeek and GLM 5.2 tie on code but sit 13 points apart on design. Qwen3.8 Max is too new to have a design rating yet.
For most work, follow the code score. A dashboard has to be right before it has to be pretty, and a wrong number in a beautiful chart is worse than a plain one that is correct. Switch to the design score when the output is going in front of a customer, a board, or anyone you are trying to impress.
Reading the three scores
All three run from 0 to 100. Higher is better. These are automated test scores, so read them as a tiebreaker between models that human voters already rank about the same — not as the deciding factor on their own.
| Score | What it measures | When it matters to you |
|---|---|---|
| Intelligence | General reasoning and multi-step problem solving. | Ambiguous questions, messy data, anything where the right answer depends on understanding your business. |
| Coding | Writing, debugging, and refactoring code. | The main one for dashboard work. Everything you build is code underneath. |
| Agentic | Long chains of tool calls and self-correction when a step fails. | Multi-hour app builds, integrations with several moving parts, anything you leave running. |
What a point of coding costs
Two models with almost the same coding score can differ by 100× in price. So we divide each model's blended price by its coding score and publish the result: cost of intelligence, the dollars you pay per point of coding ability. Shorter bar, better deal.
Frontier costs about 99× what the budget model costs per point — for roughly 13% more coding ability. That gap is the whole argument for not running everything on the most expensive option.
Cheapest per point is not automatically the right choice. This number comes from an automated benchmark, and benchmarks reward code that passes tests rather than screens people want to look at. When real people vote on the results, the budget model drops well down the field on visual work — even though it ties the default on benchmark scores. That is why the default is a slightly pricier model. Use the cheap one for bulk and repetitive jobs, not for the dashboard someone presents to the board.
Which model for which job
Prices and scores shift as vendors ship new versions, so treat the catalog as the source of truth and this table as the decision rule.
| Situation | Reach for | Why |
|---|---|---|
| Everyday dashboards, edits, "make the chart monthly" | The Default model in the catalog | Strong coding at a fraction of frontier price. This is where most work belongs. |
| Exploring, drafting, learning the tool | A Free model | Costs nothing. Good enough to prototype an idea before you commit to building it properly. |
| A build that has failed twice already | A Frontier model | Highest quality on hard, multi-step problems. Switch back down once it is unstuck. |
| A dashboard going in front of a client or the board | Highest Design score you will pay for | Presentation-grade output is judged by eye, and the design ranking is not the same as the code ranking. |
| An overnight app build with many steps | Highest Agentic score you will pay for | Long chains fail on the weakest link. Agentic score is the one that predicts recovery. |
| High-volume routine work | The best cost of intelligence you can live with | When you are running the same shape of job repeatedly, price per million tokens dominates everything else. |
See the current catalog with live scores and prices →
A more expensive model is rarely the fix for a vague prompt. When a build comes back wrong, re-read what you asked for before you reach for a bigger model. Adding the missing definition costs nothing and works more often.
Switching models
Which model your box uses is a configuration setting rather than something you choose per prompt. Your coach set it during deployment and can change it, and depending on how your box is configured you may be able to change it yourself.
Confirm with your coach Where the Rapid.Coder model is selected for your deployment, whether you can change it yourself, and whether the change applies to the whole box or per user. Ask us and we will fill in the exact steps for your setup.
Either way, ask for the change in plain English and your coach will confirm it took effect. If you want to check what is running right now, ask the box:
Which AI model are you currently running as, and where is that configured on this box? Show me the setting and tell me whether changing it affects just me or everyone who connects.
What it costs
AI models bill by the token, which is roughly three quarters of a word. Every model in the catalog lists two prices per million tokens: input for what the model reads, and output for what it writes. Output is always the more expensive half. Add the two together and you get the blended price behind the cost of intelligence figures above.
Rough shape of a normal week, so the numbers feel real:
- A single dashboard edit reads a bit of context and writes a small change. Fractions of a cent on a mid-tier model.
- Building a dashboard from scratch means exploring your data and writing a page of code. Cents, not dollars.
- An overnight app build runs hundreds of steps and reads a lot of context. This is where the choice of model shows up on a bill.
Confirm with your coach How AI usage is billed on your agreement, whether there is a monthly allowance, and where you can see usage to date. This varies by deployment mode and contract.
Habits that save money
These matter more than model choice for most teams.
Start a fresh conversation when you change topics
Every message carries the whole conversation along with it. A chat that started with payroll and drifted into inventory is paying to re-read payroll on every turn.
Say the whole thing once
One well-specified request beats six clarifying rounds, both in quality and in cost. Say what you want, who reads it, and what finished looks like.
Write down definitions once, in the project
Ask the box to record your fiscal calendar, your revenue definition, and your exclusions in the workspace. Every future build reads them for free instead of asking you again.
Create a file in this workspace that records our business definitions, and read it at the start of every future build. Include: - our fiscal year (starts <month>) and how we label quarters - what counts as revenue and what is excluded (<list your exclusions>) - our customer segments and how a record is assigned to one - which date field is authoritative for "when a sale happened" - any entities or locations that should be excluded from company totals Ask me about anything you cannot determine from the connected data rather than guessing. Then show me the file so I can correct it.
Prototype cheap, build properly once
Sketch the idea on a free model. Once you know what you want, rebuild it on the default model and keep that version.
Talk to a human instead.
Every RapidDashboard client gets direct access to the team that built the box. Most questions take under ten minutes on a call.