How much does it cost to run a chatbot with GPT-5? Three real scenarios
By LW Forge — maintainer of LLM Scout · Updated July 6, 2026
"How much will a chatbot cost me?" has a frustrating official answer — "it depends on tokens" — and a useful practical one: for most products, somewhere between tens and a few thousand dollars a month, and you can predict where you'll land before writing a line of code. This article does the math in the open for three realistic scenarios on OpenAI's current GPT-5.6 family, so you can adapt the numbers to your own case.
A quick refresher first: you pay separately for input tokens (everything you send — system prompt, conversation history, the user's message) and output tokens (what the model answers), and output is billed several times higher. At list price in July 2026, the GPT-5.6 tiers sit roughly at $1/$6 per million input/output tokens for Luna, $2.50/$15 for Terra and $5/$30 for Sol. Prices change — treat these as the anchor for the reasoning, and check current rates in the OpenAI cost calculator before budgeting. If tokens are still fuzzy, start with why 1,000 words ≠ 1,000 tokens.
Scenario 1 — Customer support bot (Luna tier)
The classic: a bot answering common questions, escalating the hard ones. Say 10,000 conversations a month, eight exchanges each, and each request carrying about 1,500 input tokens (system prompt + accumulated history) and producing 200 output tokens.
That's 80,000 requests: roughly 120M input and 16M output tokens per month. At Luna list rates the math reads 120 × $1 + 16 × $6 — a bill in the low hundreds of dollars per month. Per conversation, that's around two cents. For a bot deflecting tickets that would cost dollars each to handle by a human, the economics are not subtle.
Scenario 2 — Internal assistant over your docs (Terra tier)
A mid-size company assistant that answers with reasoning over retrieved documents needs the workhorse tier. Say 50,000 requests a month, each carrying a heavier ~2,000 input tokens (instructions + retrieved context) and 300 output tokens: about 100M input and 15M output tokens. At Terra rates the same arithmetic lands in the mid hundreds of dollars per month — input and output contributing roughly equal halves, which is typical of retrieval-augmented workloads.
Scenario 3 — Complex agent (Sol tier)
Agents are where bills surprise people, for a structural reason: one user task fans out into many model calls. An agent that plans, calls tools and verifies might make ten model calls per task, each carrying 8,000 input tokens of instructions, tool schemas and intermediate state. At 5,000 tasks a month that's 50,000 calls — 400M input and maybe 50M output tokens. On Sol, the flagship tier, you're now in the thousands of dollars per month. The lever that matters here isn't traffic, it's calls per task and state size per call — halving either halves the bill.
The history trap
Chat APIs are stateless: every turn resends the whole conversation. Turn 2 carries turns 0-1; turn 10 carries everything before it. Input cost per conversation therefore grows roughly with the square of its length. Two consequences: long conversations are disproportionately expensive, and trimming or summarizing history after a few turns is the single highest-impact optimization in most chat products. Cap history at the last N turns plus a rolling summary and the quadratic curve flattens back to linear.
Caching changes the picture — automatically
OpenAI applies cached-input pricing to repeated prompt prefixes: cache reads are billed at about a tenth of the normal input rate. In a chatbot, the system prompt and the older history are exactly a repeated prefix — so in steady traffic much of the input in scenarios 1 and 2 gets the discount without any code changes. A large system prompt hurts far less than the naive math suggests; sparse traffic (where cache entries expire between requests) benefits least. Estimate at full price, then treat the cache as margin of safety.
Cost by volume, at a glance
For the support-bot profile of scenario 1 (Luna, ~1,700 tokens per exchange), the monthly bill scales close to linearly with volume: around a couple of dollars for 1,000 conversations, low hundreds for 10,000, and low thousands for 100,000 — before caching discounts. The pattern to remember: model tier moves the bill by multiples (Terra ≈ 2.5×, Sol ≈ 5× the input rate of Luna), volume moves it linearly, and conversation length moves it quadratically. Choose the tier per task, not per company: many products run Luna for triage and only escalate hard queries to Terra or Sol, cutting the blended rate dramatically.
Run your own numbers
Every scenario above is three inputs: tokens in, tokens out, requests per month. Plug your real values into the OpenAI cost calculator to get live per-model figures, and paste a real conversation into the token calculator to measure your actual tokens per exchange instead of guessing. The estimate you can defend is the one built from your own traffic shape.