How Long Will an AI Response Be — and What Does It Cost?

Every AI response is billed in output tokens, and output tokens are the expensive ones — typically 3–6× the input rate on every major model. A 500-word answer is roughly 665 tokens; at GPT-5.6 Terra's $15 per million output rate that's $0.01 per response, which sounds trivial until you multiply by 100,000 responses a month and get $1,000 — for one feature. This estimator converts any target length (words, pages, or tokens) into token counts, per-response cost, monthly spend, and the max_tokens value you should actually set.

The Conversion Math

The 1.33 tokens-per-word ratio holds for typical English prose — the same rule behind our tokens-to-words converter. Code, tables, non-English text, and heavy formatting tokenize less efficiently (1.5–2+ tokens per word), so pad estimates for those workloads. The 25% headroom on max_tokens prevents mid-sentence truncation while still capping runaway responses.

Output Pricing Across Models (July 2026)

• Claude Fable 5: $50/M output — the premium frontier tier
• GPT-5.6 Sol: $30/M · Terra: $15/M · Luna: $6/M
• Gemini 3.1 Pro: $12/M · 3.5 Flash: $9/M · 3 Flash: $3/M · Flash-Lite: $0.40/M
• Claude Sonnet 5: $10/M · Haiku 4.5: $5/M
• DeepSeek V4 Flash: $0.28/M — roughly 178× cheaper per output token than the priciest frontier model

Confirm live rates before budgeting — provider pricing pages: OpenAI, Anthropic, and Google. Models move price quarterly in 2026's market.

How to Use This Estimator

Enter your target response length — 500 words for a support answer, 2 pages for a report section, or a straight token figure if you think in tokens — then pick the model and, optionally, your monthly response volume. You'll get the output token estimate, per-response and monthly cost, reading time, and a suggested max_tokens cap. Then flip between models: seeing the identical workload at $50/M versus $0.28/M is usually the moment teams adopt model routing.

Worked Example

A chatbot answers 10,000 questions monthly at ~500 words each (665 output tokens). On GPT-5.6 Terra ($15/M): $0.00998 per response, ~$100/month. Routing the 70% of simple questions to Gemini 3 Flash ($3/M ≈ $20/month for that share) and keeping Terra for the hard 30% (~$30) lands near $50/month — half the bill, no visible quality drop. Add a max_tokens cap of ~830 and verbose answers stop inflating it further. Output cost is a design decision, not a fixed fee.

Why Responses Cost More Than You Expect

• Output is the expensive direction: $15/M output vs $2.50/M input on GPT-5.6 Terra — a 6× gap; verbose responses hurt disproportionately
• Reasoning tokens are invisible output: thinking-mode models bill their hidden chain-of-thought at output rates — reasoning-heavy calls can cost 2–5× the visible text
• Models elaborate by default: uncapped and unprompted, responses routinely run 2–3× longer than the task needs — "be concise" plus a max_tokens cap is a 30–50% cost cut
• Defaults are oversized: a 4K-token default buffer against a 200-token task is a 20× overshoot waiting for a verbose day

Sizing the Rest of Your AI Bill

Output is one side of the ledger. Model total request costs — input, context, and caching included — with the LLM API cost calculator, count real prompt sizes with the AI token counter, and budget retrieval pipelines with the embedding cost calculator. If these responses power a paid product, run the per-customer total through the SaaS pricing calculator — AI output is a cost to serve, and it belongs inside your price, not underneath it.

Frequently Asked Questions