AI cost calculator: BYOK vs subscriptions.

How the calculator works

The estimates use these per-minute averages, derived from real conversation logs across founder workflows:

• Input tokens per minute: ~4,000 (you typing, plus the AI re-reading prior context)
• Output tokens per minute: ~1,600 (the AI's response stream)
• Working days per month: 22

Pricing is per the providers' published rates as of early 2026. Numbers are approximate; your actual usage will vary based on conversation length, context window size, and model behavior.

What v2.0 changes for the math

Keepance v2.0 added a few things that nudge real-world API spend in different directions. Worth knowing before you set your cap.

• Image attachments add vision tokens. Drop a screenshot or a whiteboard photo into chat, and Claude / OpenAI / Gemini bill it as image tokens, which are pricier per image than the equivalent text. The chat input shows a token-cost meter so you can actually see this before you send. Details in the v2.0 announce post.
• PDF chat and PDF RAG are billed as text. Native PDF on Claude / Opus is a per-page surcharge; the text-extraction path on other providers is just whatever tokens the extracted text takes. The "Include PDFs in workspace index" toggle is local: indexing doesn't call the API, only chat does.
• Read aloud doesn't cost a thing. The Piper TTS sidecar runs locally on your machine. No cloud API, no token meter.
• Plugins with the ai:invoke permission use your BYOK key, not Keepance's. So a plugin that calls the AI is just more of your monthly usage, billed exactly the same way as your own chats.
• Ollama as a fourth provider. If you wire up Ollama for local inference, the API cost for those conversations is $0. Honestly this is the cleanest hedge against a runaway month. Quality lags Claude / GPT / Gemini for the strategic stuff, but for volume drafting it's a free fallback.

Why BYOK usually wins

Subscription tools charge a flat $20/month regardless of usage. Most founders have lumpy AI usage: $0 some weeks, heavy others. BYOK matches the lumpiness. You pay for what you actually use.

For light users (under 30 minutes daily), BYOK with Claude Sonnet or Gemini Flash typically runs $3-12/month. For heavy users (over 2 hours daily on Claude Sonnet), BYOK can hit $40-60/month, which is more than a single subscription but typically still less than the stack of subscriptions most founders pay for.

What's not in the model

• API credits sometimes included in subscriptions (e.g., the small free tier OpenAI bundles with ChatGPT Plus)
• Volume discounts for very high usage (which generally don't apply at indie scale)
• Cost of Cursor / Continue.dev / similar for code work (which is its own bucket)
• Variable token usage based on context window (longer conversations = more tokens per turn)

How to set a hard spending cap

All three providers (Anthropic, OpenAI, Google) let you set a hard monthly cap so you can't accidentally exceed your budget. Set it to 2x your estimated monthly usage as a safety margin.

• Anthropic: Console → Plans & Billing → Spending limit
• OpenAI: Settings → Limits → Hard limit
• Google: Cloud Console → Billing → Budgets

The full setup guide is at /api-key-setup-guide.