Paste your real system prompt and a few real conversation turns, pick a real model from live API pricing, and get a real cost-per-conversation number — including the part most people miss: a multi-turn chat resends the entire conversation history on every turn, so cost grows with conversation length, not just message count.
Token counts below use the real o200k_base BPE tokenizer (via the gpt-tokenizer package) — the actual tokenizer behind OpenAI's GPT-4o family. Anthropic, Google, xAI and DeepSeek don't publish a client-side tokenizer, so this same count is used as a disclosed stand-in for all of them: exact for one model family, a close approximation for the rest since tokenization differs slightly by model. The arithmetic on top of whatever count you see — tokens × each model's published price — is exact.
Turns per conversation (6) is more than the 3 example turns above — the extra turns are extrapolated by repeating your examples' average length, not pasted text.
OpenAI
Anthropic
xAI
DeepSeek
≈ 600 conversations/day · 18,000/month, at 6 turns each.
Every turn resends the system prompt plus the entire conversation so far as input — the bar below is each turn's total input tokens (prior history in the dim segment, system prompt + new message in the solid one). For gpt-6-astra, that compounding makes this conversation cost $0.0447 — 88% more than the $0.0238 you'd get by naively multiplying turn 1's cost by 6 turns.
| Turn | Input tokens (history + new) | Output tokens | gpt-6-astra cost | Claude Fable 5.1 cost | gemini-3.8-flash cost |
|---|---|---|---|---|---|
| 1 | 184 | 106 | $0.00396 | $0.00714 | $0.000536 |
| 2 | 325 | 106 | $0.00537 | $0.00855 | $0.000641 |
| 3 | 466 | 106 | $0.00678 | $0.00996 | $0.000747 |
| 4 | 603 | 106 | $0.00815 | $0.0113 | $0.00085 |
| 5 | 740 | 106 | $0.00952 | $0.0127 | $0.000953 |
| 6 | 877 | 106 | $0.0109 | $0.0141 | $0.001055 |
Dim segment = every prior turn's user and assistant tokens, resent as conversation history. Solid segment = the system prompt plus this turn's new user message (the system prompt is resent in full on every turn too, not just the first). Bars are scaled to the largest turn (877 tokens, turn 6).
OpenAI · $10 in / $20 out per 1M tokens
Anthropic · $10 in / $50 out per 1M tokens
Google · $0.75 in / $3.75 out per 1M tokens
Context-window figures are a small, hand-maintained lookup table — approximate, and kept separately from the verified pricing data above (which deliberately has no context-window field; see the page source for why). Treat them as directional, not authoritative.