Method & assumptions
Everything the calculators assume, in one place. Correct us and we'll fix it.
Where model prices come from
Per-token model prices are fetched daily, automatically from OpenRouter's public pricing API (/api/v1/models), converted to USD per 1M tokens, and rendered into calculators unchanged. The raw snapshot is diffed against the previous day; moves appear on the changes page. A second public source (LiteLLM's community price list) is fetched as a cross-check; divergences greater than 1% are flagged internally. If a day's cross-check fails, the table is still published with the primary source and date stamped.
Embedding models are a different case: the main model API does not list embedding endpoints at all, so the embedding prices in the price table and in the RAG / embedding calculators come from the second source (LiteLLM). If that fetch fails, the previous day's embedding list is reused unchanged and the date stamp shows it; if no list is available at all, those two calculators fall back to a manual $/1M rate input rather than silently assuming zero.
Which models appear in the calculators
Two deliberate exclusions, both visible as filters on the pricing table:
- Models that return images, audio or video are listed in the price table but kept out of every calculator dropdown. They bill on a different axis (per image, per audio second) and comparing them per token produces numbers that look plausible and are wrong.
- Batch and priority tiers (:batch variants and similar) are excluded, consistent with the "batch discounts deliberately excluded" rule below. Selecting one would quietly produce a discounted estimate while the page claims a conservative one.
Non-token rates (editable inputs)
Rates that aren't per-token are user-editable defaults. Each is marked with its source family and "as of" date on the calculator page. We deliberately do not hardcode vendor prices we can't re-verify daily — you should paste the number from your actual vendor invoice.
| Rate | Default | Basis |
|---|---|---|
| Speech-to-text | $0.005 / min | Range of mainstream hosted STT APIs (Whisper-class to Nova-class), 2026 |
| Text-to-speech | $0.015 / min | Mid-tier neural TTS, ~450 chars/min speech |
| Telephony (inbound) | $0.015 / min | Typical US SIP/voice-API inbound rate, excl. number rental |
| Email verification | $0.01 / lead | Volume tier of verification APIs |
| Vector database | $50 / month | Small-to-mid managed index; edit to your vendor |
| Words → tokens | ×1.33 | ≈0.75 words per token for English (BPE averages) |
What we deliberately exclude
- Prompt caching discounts — real but workload-specific; results without caching are the conservative case.
- Batch API discounts — same reason.
- Self-hosted GPU math — different problem class; these calculators are for hosted APIs.
The one number that matters
Industry breakdowns of production AI systems consistently find that LLM inference accounts for roughly 15–35% of total operating cost — the rest is speech, data infrastructure, retries, observability, and human review. That's why the Total Cost calculator exists: it shows the split, not just the token bill.
How accurate is an estimate from this site?
Accurate enough to compare designs and to find the dominant cost line; not accurate enough to be a quote. Three sources of error are worth stating explicitly.
- Word-to-token conversion. The calculators use 1.33 tokens per English word, which is a reasonable average across modern tokenizers but varies with vocabulary, formatting, code and language. Real workloads commonly land between 1.25 and 1.45, and non-English text can be materially higher.
- Editable non-token rates. Speech, telephony, verification and hosting defaults are stated ranges, not your contract. Every one of them is an input precisely so you can replace it with the rate you are actually billed.
- Deliberate conservatism. Caching, batch discounts and barge-in are all excluded. Each of them lowers real bills, which means estimates here tend to sit slightly above what a well-optimised system actually pays.
Treat the output as a structure: which lines exist, roughly how large they are relative to each other, and which lever moves which line. That is the durable part, and it survives your vendor changing a price.
Corrections policy
Prices are fetched daily and never edited by hand, so a wrong number in the table means a wrong number in the upstream source — tell us and we will check it against the vendor's own pricing page. Formulas, defaults and explanatory text are human-written and can be wrong in ordinary ways; corrections are applied to the page and recorded on the changes page rather than quietly fixed, so anyone who relied on an earlier figure can see what moved.