# Context creep cost, the increase nobody shipped on purpose > Context creep is the slow growth of prompt size across releases. It raises cost per call with no traffic change. How to measure and cap it. URL: https://getculpa.com/context-creep-cost Last reviewed: 2026-08-01 ## Answer Context creep is the slow growth of your prompt across releases, as instructions, examples and history get added and nothing is removed. It raises cost per call with no traffic change at all. Culpa, a local-first LLM cost, margin, and forecast ledger, tracks tokens per call over time and maps a rise to the release that caused it. ## Why this happens Nobody ships context creep deliberately. A prompt gains one more instruction to fix an edge case, then two few-shot examples, then a longer history window because a user complained about forgetfulness. Each change is small and each is justified. The sum is a system prompt that doubled over a quarter, billed on every single call. It's invisible because no individual change is large enough to notice on the bill. ## What this usually looks like - Tokens per call trend upward across releases while call volume stays flat. - Your oldest and most-edited feature has the highest cost per call, and nobody set out to make it that way. - Nobody can say how long the current system prompt is without going and counting it. - Cost per user rises quarter over quarter with no pricing or traffic change to explain it. ## Common mistakes - Adding instructions without removing any. Why it hurts: Prompts only grow under this policy, and every added token bills on every call forever after. Do instead: Treat the prompt as a budget. New instructions have to displace old ones, or justify their cost. - Fixing a rare edge case by editing the shared system prompt. Why it hurts: You pay for that fix on all traffic to serve a fraction of it, which is the worst possible trade. Do instead: Handle rare cases in a branch that only rare cases hit, or in a cheap second pass. - Sending the whole conversation history on every turn by default. Why it hurts: Input grows linearly with turns, so a long session pays for its own opening message many times. Do instead: Use a rolling window or a summary, and measure the quality difference before assuming you need everything. ## Self-check - Count the tokens in your current system prompt. Most teams are surprised by the number. - Find the same prompt three months ago in version control and count it again. The delta is your creep rate. - Multiply the delta by your monthly call count and your input rate. That's what creep has cost you. - Check what share of each call's input is the system prompt rather than the user's actual request. - Read the prompt end to end and mark every instruction that exists for an edge case. Price that block. ## A system prompt that grew 900 tokens over a quarter (illustrative) A feature on GPT-5.4 at a real $0.0025 per 1k input rate, serving 3 million calls a month. The system prompt went from 1,100 tokens to 2,000 across a quarter of small edits, none of which looked expensive. Before: 3,000,000 calls x 1.1k tokens = 3,300,000k tokens x $0.0025 = $8,250 After: 3,000,000 calls x 2.0k tokens = 6,000,000k tokens x $0.0025 = $15,000 Increase = $6,750 a month, from 900 tokens nobody thought twice about Cost of a single added token at this volume = 3,000 x $0.0025 / 1,000 = $7.50 a month So the 40-token instruction added to fix one edge case costs about $300 a month At this volume every token in the system prompt costs about $7.50 a month, forever. A forty-token edge-case instruction is a $300 monthly line item that never appeared in any review. ## Cost figures Every figure carries its confidence and its source. No figure on this site is provider-reported. - $7.50 per token per month — cost of one extra system-prompt token at three million calls a month [calculated] Source: 3,000,000 calls x $0.0025 per 1k input tokens / 1,000, using the GPT-5.4 rate from the price book, effective 2026-07-02. - $8,250 to $15,000 — modelled monthly input spend before and after a 900-token prompt increase [estimated] Source: Both endpoints from the teardown arithmetic at the real GPT-5.4 input rate. A range because the call volume is modelled. ## FAQ Q: How much does one extra token in a system prompt cost? A: Multiply your monthly call count by your input rate per token. At three million calls on a $0.0025 per 1k rate, a single token costs about $7.50 a month, and it keeps costing that until somebody removes it. Q: Is prompt caching a fix for context creep? A: It reduces the rate you pay on the repeated block, which helps. It doesn't stop the block growing, and it does nothing if the growth is in conversation history rather than the fixed prefix. Caching treats the symptom. Q: How do I catch context creep before it compounds? A: Track tokens per call as a metric with an alert, the same way you would track latency. Creep is undetectable per release and obvious across a quarter, so the measurement has to be continuous rather than occasional. ## Sources - OpenAI API pricing: https://developers.openai.com/api/docs/pricing Run the free Cost Leak Scan: https://app.getculpa.com/scan?source=pseo&slug=context-creep-cost&cluster=problem Machine-readable index of every guide: https://getculpa.com/api/pages Human-readable index of every guide: https://getculpa.com/guides Site overview: https://app.getculpa.com/llms.txt Privacy: Culpa runs on your infrastructure. Your prompts and responses never leave it. Culpa counts calls to run your plan, and it fails open, so if it ever breaks your app keeps running.