--- title: "How do you reduce token costs without creating more rework?" description: "A smaller token bill can hide extra retries and review. Compare one change using the same tasks, quality criteria and cost per accepted result." canonical: https://www.minglabs.com/insights/answers/how-do-you-reduce-token-costs-without-more-rework lang: en last-updated: 2026-10-18 --- # How do you reduce token costs without creating more rework? [Home](/) [Insights](/insights) [Answers](/insights/answers) How do you reduce token costs without creating more rework? [English](/insights/answers/how-do-you-reduce-token-costs-without-more-rework)[Deutsch](/de/insights/answers/wie-senken-sie-tokenkosten-ohne-mehr-nacharbeit) Answer Measure the whole request before changing the model or shortening its context. Compare one change on the same tasks, with the same acceptance criteria, and include retries and human correction. Keep the saving only if quality still meets the agreed bar and cost per accepted result falls. Our worked example shows why a smaller model bill can still leave the team with more work.[\[S1\]](#source-s1) Invented example · the same 100 requests The same €3 token saving can lead to different total costs. Model usageReview and correction Baseline**€68.00** Model usage: €8.00 · Review and correction: €60.00 · Accepted drafts: 90/100 Less repeated context**€65.00** Model usage: €5.00 · Review and correction: €60.00 · Accepted drafts: 90/100 Too much context removed**€95.00** Model usage: €5.00 · Review and correction: €90.00 · Accepted drafts: 85/100 Chosen assumptions: model usage falls from €8 to €5 in both alternatives. With unchanged review, combined cost falls from €68 to €65. With more rework, it rises to €95 and fewer drafts are accepted. These are arithmetic scenarios, not observed outcomes. Only model usage and review are counted. This is not the full operating cost. Last updated: October 2026 | Next review: November 2026 Worked example and proposed method [Machine-readable record ](/insights/answers/how-do-you-reduce-token-costs-without-more-rework.json) By MING Labs · Editorial team The answer at a glance 1. [Follow one request through every model call.](#request) 2. [Remove work only when the task still succeeds.](#changes) 3. [Include the correction work in the comparison.](#comparison) ## Follow the customer request, not just the first model call. Imagine an agent drafts replies about delivery dates. It reads the customer’s question, looks up an order and prepares a reply for a colleague to approve. When the order number is missing, it should ask for it. If you compare only the first model call, you miss the second lookup, another attempt at the answer and the colleague who repairs it. We would start with a record for each request: the task, model and settings, all usage and tool charges, retries, result, review time and acceptance decision. Tokens are the units used to process text; input and output can have different rates. Read the provider’s billing fields and reconcile them with the actual charge. A subscription allowance needs its own accounting; a token count is not automatically an extra invoice. Define a usable reply before testing. In this example it uses the right order, gives only a supported date and asks for missing information. Record a correct handoff separately from a completed reply. The same definitions must apply to both versions.[\[S1\]](#source-s1) ## Which change is worth testing on your task? Discuss how to examine usage and the correction work together with MING. [Discuss my token costs](/contact?type=project&topic=token-cost-review) Tell us what you have in mind. We will respond personally. ## Remove repeated work before taking away information the task needs. A useful first experiment is to return the order status and expected date instead of the entire order history. Another is to remove a repeated instruction or ask for the specific customer reply rather than a long internal summary. Change one thing so you can explain the result. OpenAI’s optimization guide follows a baseline, change and quality-check sequence; its simulated savings are not a forecast for your workload.[\[S2\]](#source-s2) Keep the order’s identity, relevant exception and source date. If the shorter record hides a split shipment, the agent may produce a confident but unusable answer. The person approving the reply should see the same evidence and use the same criteria in both versions. Compare routine requests and important exceptions separately, so a better average cannot conceal a new serious mistake. There is a concrete reason to test this. In an April 2026 postmortem, Anthropic reported reverting a reduced reasoning default and a verbosity instruction after quality problems. It also fixed a bug that repeatedly removed earlier reasoning. These were product-layer issues; the report says the API was unaffected. It is a counterexample to assuming that less thinking or less text is always an improvement.[\[S4\]](#source-s4) ## Caching helps repeated input; check that reuse actually happens. If many requests begin with the same instructions, prompt caching can reuse that prefix. Put stable material together and keep the changing order details separate where the provider’s interface permits it. Anthropic’s documentation distinguishes cache writes, reads and uncached input. A changed prefix or expired cache can prevent reuse; output still has to be generated.[\[S3\]](#source-s3) For the delivery reply, inspect a cold request and later repeated requests. Compare billed totals, including any cache-write charge. Do not pad a short prompt merely to reach a cache threshold. If repeated lookups, retries or human correction dominate the cost, a cache discount may address only a small part of the problem. ## A €3 model saving is lost when correction adds €30. Our invented example compares the same 100 initial requests. We choose rates of €2 per million input tokens, €8 per million output tokens and €60 per review hour. These are arithmetic assumptions, not current supplier prices or a MING offer. All token totals include additional attempts; do not add the retry count to the bill again.[\[S1\]](#source-s1) The example keeps the starting workload and assumed rates fixed; every outcome below is invented. | For 100 initial requests | Baseline | Less repeated context | Too much context removed | | --- | --- | --- | --- | | Input tokens (millions) | 2 | 1.5 | 1 | | Output tokens (millions) | 0.5 | 0.25 | 0.375 | | Additional attempts | 20 | 12 | 25 | | Model usage | €8.00 | €5.00 | €5.00 | | Review and correction (minutes) | 60 | 60 | 90 | | Review and correction | €60.00 | €60.00 | €90.00 | | Combined cost | €68.00 | €65.00 | €95.00 | | Accepted drafts | 90 | 90 | 85 | | Combined cost per accepted draft | €0.76 | €0.72 | €1.12 | Model cost is input millions × €2 plus output millions × €8. Review cost is minutes ÷ 60 × €60. Divide their sum by accepted drafts: €68 ÷ 90 is about €0.76; €65 ÷ 90 is €0.72; €95 ÷ 85 is €1.12. The last variant has the same €5 model bill as the leaner one, but its assumed extra correction makes each accepted draft more expensive. These scenarios show how to calculate the decision; they do not show that removing context caused these outcomes in a real test. Nor does an equal acceptance count establish equal quality: inspect the individual errors and any non-negotiable requirements. The full monthly budget also includes costs outside this example, covered in our [answer about operating an AI agent](/insights/answers/what-does-it-cost-to-run-an-ai-agent-in-production) . ## Keep the change only after reviewing the failures as well as the bill. Run the current version and the proposed version on the same representative tasks under comparable conditions. Record the model version, settings and source records so another person can repeat the comparison. Where responses vary, use repeated runs and a held-back set; a single successful demonstration is not evidence that the failure rate has stayed the same. Before the comparison, agree which failures block release and what variation in quality, review time and response time is acceptable for this task. Keep unauthorised actions and invented delivery promises out of the savings calculation: a cheaper failure is still a failure. Our [quality-measurement answer](/insights/answers/which-metrics-show-whether-an-ai-agent-is-improving) explains why correct completion and correct escalation need separate counts. Evidence and limits - The requests, rates, token totals, retries, review times and acceptance counts are invented. No model benchmark or customer saving was measured. - The proposed method combines public technical guidance with our own arithmetic. It offers no universal quality threshold, sample size or guaranteed saving. - Only model usage and review are included. Tool fees, implementation, hosting and ongoing support are excluded. Zero accepted results make unit cost undefined. Choose one repeated task and one change worth testing. Save its current requests, agreed checks and complete cost record before editing the prompt. That gives the team a defensible reason to keep the change, revise it or restore the previous version. Citable short version To reduce token costs, record billed usage across the complete task, including retries. Test one change against the same representative tasks and predetermined quality checks. Count review and correction alongside model cost, then divide by accepted results. MING proposes this acceptance method; its worked figures are invented assumptions, not measured savings or a guarantee that quality will remain unchanged. Invented inputs and outcomes; official technical sources and a provider postmortem. No measured savings or customer benchmark. ## Where does the work in your customer reply arise? Describe a repeated task and the change you want to test. We can discuss which usage, quality and review records would make the comparison useful. [Discuss my token costs](/contact?type=project&topic=token-cost-review) Tell us what you have in mind. We will respond personally. [Understand ongoing operating costs ](/insights/answers/what-does-it-cost-to-run-an-ai-agent-in-production) ## Sources \[S1\] Token-cost acceptance example MING Labs · 18 October 2026 Supports: €8/€5/€5 model usage; €60/€60/€90 review; €68/€65/€95 combined; 90/90/85 assumed accepted drafts, Retries included in total usage; zero accepted results yield undefined unit cost Original invented arithmetic comparison and proposed review method; 100 assumed requests, no live model calls or client outcomes. Chosen rates are not supplier prices. \[S2\] [Optimizing customer support agents for cost and quality](https://developers.openai.com/cookbook/examples/agent_optimization/optimizing_agents_for_cost_and_quality) OpenAI · Read 18 October 2026; reference table verified 14 September 2026 Supports: Baseline, one change, quality check; inspect prompts, output, tools and context before accepting a saving Official worked guide; synthetic tickets and deterministic simulation are not production results. Optional live code was not executed. \[S3\] [Prompt caching](https://platform.claude.com/docs/en/build-with-claude/prompt-caching) Anthropic · Read 18 October 2026 Supports: Reusable identical prefixes; separate cache write/read accounting and cache lifetime; no removal of output generation Current technical documentation; no independent savings benchmark. \[S4\] [An update on recent Claude Code quality reports](https://www.anthropic.com/engineering/april-23-postmortem) Anthropic · 23 April 2026; read 18 October 2026 Supports: Reduced reasoning default and a verbosity instruction were reverted; context-management bug was fixed, Counterexample to assuming less context or shorter answers preserve quality Provider postmortem on product changes; not an API-wide regression or universal effect size. ## Frequently asked questions Should we switch to a smaller model first? That is one candidate change. First establish which part of the task consumes the budget and which errors would make the result unusable. A lower token price does not establish that the model can perform your task. Does prompt caching mean fewer tokens are generated? No. It reuses eligible repeated input prefixes. Output still has to be generated. Check actual cache reads, writes, misses and the applicable rates before claiming a saving. How many test requests are enough? There is no universal number. Cover the task types, languages and important exceptions you actually handle. The 100 requests on this page are chosen arithmetic inputs, not a recommended sample size or a statistical quality guarantee. What if no result is accepted? Cost per accepted result is undefined. Report the spend, zero accepted results and the failures; do not display a zero cost per successful result. Are these the full operating costs of an agent? No. The example includes only model usage and review or correction. Implementation, tool fees, hosting, ongoing support and other costs require their own accounting. ## Available actions - **Copy answer for my AI**: copies the [AI briefing](https://www.minglabs.com/insights/answers/how-do-you-reduce-token-costs-without-more-rework.md) to the clipboard in the browser. --- Canonical: https://www.minglabs.com/insights/answers/how-do-you-reduce-token-costs-without-more-rework Machine-readable index: https://www.minglabs.com/llms.txt