How the math works

The formula

• A month is 30 days. Inference bills scale with usage, not calendar quirks.
• "Calls" are individual LLM API requests. One agent run that makes 8 model calls counts as 8, not 1 — this is the most common way people underestimate agent costs.
• Token counts are point estimates. If your prompt size varies a lot, enter a usage-weighted average.

Prompt caching

Most providers bill input tokens served from their prompt cache at roughly 10% of the fresh rate. The cache hit rate slider sets the share of input tokens billed at the model's published cached rate; models without published cache pricing (most open-weight routing) bill everything at the full input rate, and the calculator tells you when that happens. Real-world hit rates: a chat app with a fat system prompt easily reaches 60–80%; RAG with ever-changing retrieved chunks might see 20–30%.

Where the preset numbers come from

Presets are opinionated starting points for common AI app shapes — every value stays editable.

• Chat assistant: 1–3 short sessions a day. Input is mostly a reused system prompt plus growing history, so cache hit rates are high.
• RAG / knowledge base: Retrieved chunks inflate the prompt to several thousand tokens and change per query, so caching helps less.
• AI agent / automation: A single agent run is many LLM calls with a large, mostly-repeated context. Count steps, not runs — 2–3 runs a day is easily 20 calls.
• Writing copilot: Frequent small calls while the user works. Output-heavy relative to input, which matters because output tokens cost 4–6x more.
• Summarization pipeline: Few calls, huge inputs, small outputs. Input price dominates, so cheap long-context models win. Batch APIs can halve this again.
• Classification / extraction: Very high call volume, tiny payloads. Nano-tier models are usually indistinguishable from flagships on this job.

Pricing rules

• All rates are USD per 1M tokens on the standard synchronous API tier, verified July 2026 against public vendor pricing pages (linked on themodel prices page).
• Where a provider has context-length price tiers (e.g. Gemini 3.1 Pro above 200K tokens, GPT-5.5 above 272K), we use the base tier — the calculator's per-call token counts stay well below those thresholds.
• Claude Sonnet 5 is shown at its standard $3/$15 rate, not the intro pricing that expires August 31, 2026.
• Workers AI is billed in neurons ($0.011 per 1,000); we use Cloudflare's published per-token equivalents and ignore the 10k free daily neurons.
• OpenRouter rates are the listed default routing for each model and can vary with the underlying provider; OpenRouter also adds a ~5.5% fee when buying credits.

What we deliberately don't model

• Batch APIs — most providers cut ~50% for async workloads. If your pipeline can wait, halve the summarization-style numbers.
• Negotiated / committed-use discounts — enterprise deals commonly take 20–50% off list price.
• Reasoning token overhead — reasoning models bill their thinking as output tokens. If you enable high reasoning effort, raise output tokens per call accordingly.
• Free tiers and rate limits — they matter at prototype scale, not at the 1k–100k DAU range this tool targets.
• Everything that isn't inference — hosting, vector databases, embeddings, observability. Inference dominates AI app costs at scale, and keeping the tool focused keeps it honest.

Sanity-checking a result

Take the chat preset at 10k DAU on GPT-5.4 mini: 10,000 users × 8 calls × 30 days = 2.4M calls. Input: 2.4M × 1,200 = 2.88B tokens; with a 60% cache rate, 1.15B tokens bill at $0.75/M ($864) and 1.73B at $0.075/M ($130). Output: 2.4M × 400 = 960M tokens at $4.50/M ($4,320). Total ≈ $5,314/month — about $0.53 per user. If a model's answer surprises you, run this arithmetic; it should always reproduce the calculator's number.