AI Cost Navigator
Size your AI workload, we'll guide you through
Describe the use case, size the tokens, weigh what matters, pick a model and see cost per outcome — low / base / high — with the tuning levers that decide what a successful outcome really costs.
Describe the use case
Parameters
Priorities
What matters for this workload. Starts from the archetype's usual profile — override anything. Every weighted category below drives which model gets priced (step 3), not just the table's sort order.Priority preset
Weights
Model & volume
Size the tokens, then the best fit for this workload is ranked and priced automatically below — click any row to price that model instead.Token profile — low / base / high vary tokens-per-word, retries and cache hit. Click one to make it the active scenario — it drives everything below.
Filters
Hard floor from step 2's priorities — can't go lower. Slide up to require a smarter (usually pricier) model.
Ceiling at today's token volume. Leave blank for no ceiling.
Fit score vs. monthly cost (chart)
Top-left is the sweet spot. Cost axis is logarithmic.
How the fit score works
- Each metric is scaled to 0–100 across all models that report it. Indices are scaled linearly; price, cost per task, speed and latency span orders of magnitude and are scaled on a log basis.
- Fit score = weighted average of the scaled metrics. With "only complete data" off, missing metrics are skipped and remaining weights re-normalized (marked "partial").
- Monthly cost here uses this use case's own sized tokens above, not a generic workload.
Or set a price manually
Business case
Sizing → model → what it's worth. Model + embeddings only — infra via the Azure estimator, labour via the ledger.Model
Bill of materials (base scenario)
| Component | Role | SKU hint | Quantity | Unit | Source |
|---|
Business case — before vs. after the optimization levers below
Before tuning vs. the manual baseline from step 1
After tuning with the levers below applied
Optimization levers — what-ifs on top of the priced model above
What if more of the repeated prompt came from cache?
What if each call sent less history or fewer documents?
What if answers were shorter?
Agentic archetype only — stops runaway loops.
Total cost of AI includes infra, tooling and people around the model call — tokens are typically only 10–25% of it. Replace with your own figures when known. Only fill in "extra value" if AI creates value beyond replacing the manual baseline above (e.g. it does more, or faster than the business case above already counts) — otherwise leave it blank.