Interactive AI Model & Agent Tools

Five focused utilities use source-linked data or generate transparent design artifacts. Calculators expose assumptions; design tools do not claim to replace evaluation or production review.

How to choose between these tools

Each utility answers a different question, and using the wrong one produces a number that looks precise but answers nothing. Start from the decision you actually need to make rather than from the tool that sounds most impressive.

• You have a workload and need a budget. Use the API cost calculator. It multiplies your own call volume, token mix, and cache share against published per-million-token rates, so the output is only as good as the traffic sample you enter.
• You have a single prompt and want its size. Use the token counter. It estimates tokens for pasted text and previews one request across the verified catalog. Treat the estimate as an approximation: tokenisation differs by provider and, in Anthropic's case, changed with the Claude 4.7 tokenizer.
• You are comparing published results. Use the benchmark matrix. It deliberately carries a small comparable set rather than a long leaderboard, because vendor-reported scores often differ in harness, tool access, and inference settings.
• You do not yet know which model to evaluate. Use the shortlist wizard. It applies explicit rules about workload, unit-price priority, and context need, then shows the reasoning so you can disagree with it.
• You are designing an agent, not picking a model. Use the blueprint builder. It produces a reviewable artifact covering permissions, approvals, spending limits, and evaluation rather than a score.

Where the numbers come from

Every rate in these tools is read from a single shared catalog maintained at js/model-data.js, and each entry carries the provider documentation URL and the date it was checked. Prices are copied from the provider's own pricing or model page, never from a third-party comparison table. Where a provider publishes a promotional rate, the expiry date travels with the number. Where a provider separates peak from off-peak billing, the tool shows the off-peak rate and states the peak multiplier beside it.

Limits worth understanding before you trust an output

• A unit-price comparison is not a cost comparison. A cheaper input rate loses to a model that finishes in one call instead of five, retries less, or needs fewer agent steps.
• Long-context billing is not always proportional. Several providers charge a premium across the whole request once input crosses a threshold, rather than only on the excess. The catalog notes where this applies.
• Cache behaviour is the largest variable. A cache hit can cut input cost by an order of magnitude, but only when a stable prefix is genuinely reused. Measure your own hit rate instead of assuming one.
• Cache writes bill too. Writing to cache is not free on every provider, so a workload with a low hit rate can cost more with caching enabled than without.
• Benchmark scores are evidence about one setting. They are not predictions about your repository, language mix, or review process.
• Rates move. The catalog is re-verified on a schedule and each page shows its verification date. Check that date before quoting a figure in a budget.

These tools run entirely in your browser. Text you paste is not uploaded, and no entered value is stored or transmitted. If an output leads to a decision with real cost attached, confirm the rate against the provider's pricing page on the day you commit.