# What an AI support conversation costs, and what decides whether it pays > Anthropic puts 10,000 support conversations at about $37.00 on Haiku 4.5. That's the floor. Escalation rate is the multiplier that decides whether it pays. URL: https://getculpa.com/cost-per-support-ticket Last reviewed: 2026-08-05 Rates effective: 2026-07-02 ## Answer Anthropic publishes a worked example putting 10,000 support conversations at roughly $37.00 on Claude Haiku 4.5, about a third of a cent each. That's the floor, and escalation rate multiplies it. Culpa, a local-first LLM cost, margin, and forecast ledger, prices each conversation so your own figure sits beside the benchmark. ## Why this happens Support is the use case where AI economics look easiest and get argued about most, because the model cost is genuinely small and everything around it isn't. Anthropic's own pricing page works the example: roughly 3,700 tokens for an average conversation, priced on Haiku 4.5, giving about $37.00 per 10,000 tickets. That's a third of a cent a conversation against a human-handled ticket costing orders of magnitude more, and it makes the naive business case trivially positive. The naive case is also the one that fails review. Three things move it. Conversations that escalate cost you the model spend AND the human, so they're strictly worse than not trying. Retrieval, reranking and guardrail calls multiply the per-conversation token count well past the benchmark. And a benchmark built on a 3,700-token average says nothing about your own distribution, where the long tail of complex conversations is exactly where escalation happens. ## What this usually looks like - Your deflection business case uses a vendor benchmark rather than your own conversations. - Nobody separates conversations that resolved from conversations that escalated. - Retrieval and guardrail calls aren't counted in cost per ticket. - You know your average conversation cost and not the top decile. - Deflection rate is reported and cost per deflection never is. ## Common mistakes - Costing deflection from an average conversation. Why it hurts: The tail is where escalation lives, and the tail is much more expensive than the mean. Do instead: Price the distribution, and look hardest at the conversations that ended in a handoff. - Counting only the answering model. Why it hurts: Retrieval, reranking and safety checks are calls too, and they scale with conversation length. Do instead: Sum every call in the conversation, which is what cost per conversation means. - Treating an escalated conversation as cost-neutral. Why it hurts: You paid the model and the human, so it costs more than never attempting it. Do instead: Report cost per resolved conversation, not cost per attempted one. - Optimising the model before the flow. Why it hurts: At a third of a cent a conversation, the model is rarely the lever that matters. Do instead: Attack escalation rate and retrieval volume first, then revisit the model. ## Self-check - Work out your real average tokens per conversation and compare it to 3,700. - Split last month's conversations into resolved and escalated, and price each group. - Divide total support model spend by conversations RESOLVED, not attempted. - Count how many calls a single conversation makes, including retrieval and safety checks. ## The floor, and the two things that multiply it (illustrative) Anthropic's pricing page, read 2026-08-05, publishes a worked example for support: an average of about 3,700 tokens per conversation on Claude Haiku 4.5 at $1.00 and $5.00 per million, totalling about $37.00 per 10,000 tickets. Those figures are Anthropic's. The escalation and retrieval additions below are modelled, and they're the part a business case has to supply for itself. Anthropic's published figure: about $37.00 per 10,000 conversations which is $0.0037 per conversation, a third of a cent modelled: a retrieval step and a safety check triple the calls per conversation modelled: 20% of conversations escalate, so 10,000 attempts resolve 8,000 cost per RESOLVED conversation then rises well above the headline, before a human is paid at all The published number is real, and it's a floor rather than an estimate. It prices one answering call on an average exchange, so every real deployment adds to it: more calls per conversation, and conversations that end with a person anyway. Both multipliers come from your own traffic, which is why the benchmark settles the model question and leaves the business case open. ## Cost figures Every figure carries its confidence and its source. For each provider-reported figure below, its own source line names the published record it came from, the date that record was read, and any aggregation applied to it. - about $37.00 per 10,000 conversations, Anthropic's published worked example for support ticket processing on Claude Haiku 4.5 [provider-reported] Source: Anthropic's pricing page at platform.claude.com/docs/en/docs/about-claude/pricing, read 2026-08-05, gives a worked example of about 3,700 tokens per conversation on Claude Haiku 4.5 at $1.00 and $5.00 per million, totalling about $37.00 per 10,000 tickets, which is $0.0037 per conversation. Quoted as published, with no recalculation. The escalation and retrieval adjustments discussed alongside it are modelled. ## FAQ Q: How much does an AI support conversation cost? A: Anthropic's published example puts an average conversation at about 3,700 tokens on Claude Haiku 4.5, roughly $37.00 per 10,000 tickets, or a third of a cent each. Your own figure depends on conversation length and how many retrieval or safety calls each one makes. Q: Why is my cost per ticket higher than the benchmark? A: Usually because a conversation is more than one call. Retrieval, reranking and guardrail checks all bill, and they scale with how long the conversation runs. The benchmark prices the answering model on an average exchange, which is the floor rather than the figure. Q: What actually decides whether deflection pays? A: Escalation rate, not token price. At a third of a cent a conversation the model is almost never the binding constraint. A conversation that escalates costs you the model spend plus the human, so it's worse than not attempting it, and that ratio moves the business case far more than any model choice. Q: Should I measure cost per attempted or per resolved conversation? A: Resolved. Attempted flatters the number by spreading cost over conversations that still needed a person. Cost per resolved conversation is the figure that compares honestly against what a human-handled ticket costs you. ## Sources - Anthropic pricing: https://platform.claude.com/docs/en/docs/about-claude/pricing Run the free Cost Leak Scan: https://app.getculpa.com/scan?source=pseo&slug=cost-per-support-ticket&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.