01 · Define the baseline
What are you actually trying to build?
A useful estimate begins with a technical baseline and explicit unknowns—not a project-type price sticker.
0% specified
Capabilities in scope
02 · Delivery architecture
Who does what—and for how long?
Each line is an editable assumption. Rates are scenario inputs, not claimed market benchmarks.
| Workstream | Role | Low hours | High hours | Rate | Range | |
|---|---|---|---|---|---|---|
| Labor subtotal | 0 | 0 | $0–$0 | |||
Start from:
03 · Usage economics
What will the system consume each month?
Model costs are computed from your workload assumptions. Provider rates come from a dated reference catalog shown below; requests, tokens and caching are yours to change.
Rate desk
Where these prices come from, and when to stop trusting them
—
| Model | Input | Cached input | Output | Checked | Provenance |
|---|
What the catalog is
A dated planning snapshot of published list rates, typed into this page by hand and versioned with the site. It is not a live feed, it is not a quote, and it does not know your contract. It exists so that a number in the results has a traceable origin instead of being folklore.
How stale it can get
The catalog is only as current as its check date, shown above and on every export. Treat anything past thirty days as needing verification before you commit money, and anything past ninety days as a placeholder. This page tells you which of those you are looking at rather than making you work it out.
What moves a published rate
List price is only one input. The provider’s own pricing page prices batch processing at half the standard input and output rate, adds ten percent for data residency, quotes standard rates only for context lengths under 270K, prices cached input at roughly a tenth of fresh input, and bills built-in tools separately — web search was listed at $10.00 per thousand calls. Priority and flex processing tiers change it again. Verified against the archived OpenAI API pricing page of 16 June 2026.
What to do about it
Two things. Change the workload assumptions — requests, tokens, cached share — because they move the answer far more than the rate does. And if you have a negotiated, promotional, or newer rate, tick Use my own rate on the workload and type it in; the model, the exports, and the share link all carry your number and mark it as an override.
Always verify before procurement: OpenAI API pricing · AWS Pricing Calculator · Google Cloud Pricing Calculator. Workload costs also change with reasoning effort, modalities, region, volume tiers and negotiated terms.
Infrastructure and recurring services
| Service | Basis / assumption | Monthly low | Monthly high | |
|---|---|---|---|---|
| Runtime subtotal | $0 | $0 | ||
04 · Other build and year-one costs
What sits outside implementation labor?
Capture costs that disappear from software estimates: data, review, procurement, training, launch, and ongoing care.
One-time costs
Monthly operating costs
05 · Uncertainty & governance
Where could the estimate move?
Risk is expressed through named drivers and adjustable impacts—not an unexplained blanket markup.
Required review lanes
Checking a lane records that review is in scope; it does not certify compliance. Add its labor and external cost explicitly in the model.
06 · Decision view
The scenario, with its assumptions attached
Use this to compare options and prepare discovery—not as a substitute for a project-specific quote.
Specification confidence0%
Reading this scenario
Cost composition
Year-one contribution of each block at its high end, so the bars are comparable.
Where the money actually is
Every line item, ranked by how much of the year-one range it is responsible for.
Sensitivity — one input at a time
Each bar is the full width of a single assumption’s own range, holding everything else fixed. The longest bar is the assumption worth pinning down first; the short ones are not worth another meeting.
The arithmetic, on your numbers
Every figure above is reproducible by hand. This is the same calculation the page just performed, written out.
Scenario brief
What this number is not
- Not a quote. Nothing here has been priced by anyone who would be doing the work, and no one is bound by it.
- Not fully loaded. Labor is entered as a rate per hour. If those are employees rather than a vendor, the scenario excludes benefits, payroll tax, recruiting, equipment, software seats, workspace and management overhead — commonly a large multiple of base pay.
- Not a schedule. Hours are not calendar time. Two hundred hours does not become five weeks because a week has forty hours in it; sequencing, review cycles, and waiting on other people are not modeled.
- Not a probability. The risk drivers are percentage impacts you chose, applied to the high build subtotal. They are not likelihoods, confidence intervals, or a compliance posture.
- Not indexed. Rates, wages and vendor prices are held flat across the twelve months. Over a longer horizon that assumption is the first one to break.
What to do with it tomorrow
- Attack the widest bar, not the biggest total. The sensitivity list ranks assumptions by how much range each one contributes. The top one is the cheapest uncertainty to remove.
- Separate the estimate from the decision. Build is one-time and mostly labor you can scope. Run rate compounds every month and is mostly workload you can measure. They fail differently and should be argued separately.
- Re-derive the run rate from a real trace. Requests and tokens per request are the two inputs that dominate model cost. One week of production logs replaces the guess entirely.
- Bring the assumptions, not the total. Export the scenario or copy the share link. A reviewer who can see the labor lines and workload assumptions can argue with them; a reviewer handed a single number can only agree or disagree.
- Re-run it when the rate desk goes amber. The catalog carries its check date into the export, so a scenario from three months ago announces its own age.
Need a project-specific answer?
Bring the scenario—not just the total.
RN can review the assumptions, identify missing work, compare implementation paths, and turn the strongest option into a scoped plan.
07 · Options side by side
One scenario is an estimate. Three are a decision.
Save the scenario you are looking at, change the approach, save that too, then read the difference. Saved scenarios stay in this browser; nothing is uploaded.
| Scenario | Saved | Build | Monthly | Year one | Δ vs baseline | Spec |
|---|
How to use a comparison honestly
Two scenarios are comparable only if they were specified to the same depth. A lean prototype at 30% specification confidence will almost always look cheaper than a production build at 80% — not because it is, but because the work nobody has written down yet costs nothing. The specification column is in the table for exactly this reason: read it before you read the totals.
The useful comparison is rarely the cheapest number. It is which scenario moves a decision forward — what each one would teach you, what it forecloses, and how much of its range is uncertainty you could remove this week with a day of discovery.
Method
Transparent by design.
The workbench follows a simple principle from cost-estimation practice: define the baseline, document assumptions, model ranges, identify sensitive drivers, and update the estimate as evidence improves.
Read methodology & sources →RN
Designed and built by
RN Collins · Aloha AI
Research, science, design, and AI brought together to make complicated decisions visible and usable.