Model the whole AI workload—not one token price.
Analyze private repositories and documents locally, then project recurring API, RAG, cache, vision, agent, embedding, and infrastructure costs with every assumption visible.
Measure source code and documents before they leave your browser.
Select repository files or a directory. Supported text, source, PDF, and DOCX files are extracted and counted locally with the selected model method.
No files read yet. File contents never enter a network request.
Convert one request shape into monthly and annual spend.
Includes cached input, retry amplification, and separate input/output rates.
Separate ingestion cost from recurring retrieval and generation.
Models document indexing, embedding, retrieved context, generation, and vector infrastructure without hiding one-time costs inside a monthly total.
Find the monthly value of a reusable prefix.
Compares uncached list input with cache-read pricing. Cache-write fees and minimum lifetime rules vary by provider and remain outside this simplified comparison.
Project image tokens from geometry, detail, and batch volume.
Uses the same provider-specific tile and patch engine as the main calculator.
Expose the multiplication hidden inside agent iterations.
Each iteration can resend history, tool schemas, and tool results. This planner applies an explicit history-growth assumption rather than presenting a false exact bill.
Estimate chunk volume, overlap amplification, and re-index cost.
Enter the embedding price from the provider you plan to use; the calculator does not silently bind this result to a generation model.
Estimate weight memory and operating cost—with assumptions visible.
This is a capacity worksheet, not a throughput benchmark. Runtime frameworks, MoE activation, KV formats, batching, parallelism, and real utilization can materially change the result.
Compare the same workload across every verified catalog entry.
Prices, context limits, accuracy labels, and source links come from the shared model registry—not duplicated page copy.
Planning telemetry, not fabricated certainty.
Repository, document, text, and supported image calculations reuse TokenCalculator.dev's browser-local engines. Recurring workload tools apply the values entered above. Provider wrappers, tool serialization, cache writes, OCR, retries, taxes, gateway fees, storage operations, and measured hardware throughput can change production cost.
Use these results to compare scenarios, then reconcile representative non-sensitive traffic against provider usage metadata and invoices.
Read the calculation methodology · Review the privacy architectureFrequently asked questions
Do repository files leave my browser?+
No. Supported source, data, PDF, and DOCX files are read, extracted, and tokenized in the browser. Their contents are not uploaded for calculation.
Are these production invoices?+
No. These are planning models based on explicit workload assumptions and the verified model catalog. Provider usage metadata remains authoritative.
Why are the estimators on one page?+
They share one normalized model registry and calculation engine, preventing pricing and context-window facts from drifting between separate tools.