LEARN / CHATGPT COST PER PROMPT

How much does ChatGPT actually cost per prompt?

PUBLISHED 2026-07-10 · TOKENBURN INDEX

Somewhere between “basically free” and “more than you think.” The honest answer depends on how you pay — subscription or API — and on a multiplier almost nobody accounts for: conversation history. Here is the full arithmetic, using mid-2026 list prices, so you can stop guessing.

First: what is “a prompt” in billing terms?

Providers don't bill per question — they bill per token, the unit models read and write (roughly 4 characters of English each; see our tokenomics guide for the full primer). A realistic average exchange is ~150 tokens of input (your question) and ~450 tokens of output (the answer — models are chatty). Input and output are priced differently, and output is 4–8x more expensive, because generating text costs more compute than reading it. Every number below builds on that 150-in / 450-out baseline; if your prompts carry pasted documents or code, scale up accordingly.

The subscription math: you set the per-prompt price

A ChatGPT Plus plan costs $20/month flat (Pro tiers run $200/month, as of mid-2026 list prices). Flat rate means your cost per prompt is simply the subscription divided by how much you actually use it:

The subscription is a flat-rate buffet: the economics reward the hungry. A casual user pays nearly 8x more per prompt than a heavy user for exactly the same product — which is also why “is it worth it?” has no universal answer. Divide $20 by your real monthly usage and compare against the API table below.

The API math: the meter is visible

Pay-per-token pricing makes the cost explicit. For the baseline prompt (150 in / 450 out), at mid-2026 list prices per million tokens:

Assumes 150 input / 450 output tokens per prompt. List prices, no caching or batch discounts. Prices change — treat this as a mid-2026 snapshot.

Two things jump out. First, the spread: the same question costs 95x more on GPT-5.5 than on DeepSeek. Model choice is the single biggest lever on any AI bill. Second, the absolute numbers are small — which is exactly how they escape scrutiny until something multiplies them. Two somethings, in fact.

Multiplier #1: conversation history

Language models are stateless. The app re-sends the entire conversation as input on every turn — turn 30 re-pays for turns 1 through 29. Run the numbers on a 30-turn working session (each turn adding ~600 tokens of history): total input is ~265,000 tokens and output ~13,500. On GPT-5.5 that conversation costs ~$1.73 — an average of 5.8¢ per prompt, four times the standalone price. Keep that chat alive for days, pasting in documents as you go, and the multiplier climbs past 10–20x. This is the single biggest gap between what people estimate and what appears on the invoice, and it's why “one long chat for everything” ranks #1 in the Burn Index. The fix costs nothing: start fresh conversations, aggressively.

Multiplier #2: output length

Output tokens cost 4–8x more than input tokens. Asking GPT-5.5 for “a detailed 2,000-word analysis” (~2,700 output tokens) costs ~8¢ in output alone; asking for “the five key points, one line each” costs under 1¢ for the same substance. Verbosity is a setting, and the default is expensive. If you routinely trim the model's answers by eye, you are paying for text you delete.

Four realistic monthly bills

• — The casual user (5 prompts/day, short chats): ~$2/month via API on GPT-5.5. Plus at $20 costs 10x more — but buys the app, voice, and no meter anxiety. You're paying for the product, not the tokens.
• — The heavy user (40+/day, long working sessions): the history multiplier would push API costs past $20-60/month. The flat-rate subscription is the bargain here. This is the user OpenAI loses money on, and that's fine — it's the plan.
• — The developer script (10,000 classification calls/month): API only. On Haiku or Gemini Flash: $15–25. On GPT-5.5 out of habit: ~$140. Same output quality for this task class. Habit is expensive.
• — The agent workflow: a different beast entirely — each run is 10–40 model calls with compounding context. We gave it its own guide.

Stop estimating, measure yours

Averages are for articles. Your prompts have a real length, a real style, and a real amount of ceremony — paste one into the Burnmeter and get its token count, its cost across 12 models, and a Burn Score telling you how much of it is signal. Multiply by your monthly volume, and you have the number this article can only approximate.

Frequently asked questions