LLM cost per ticket formula: (input_tokens ÷ 1,000,000 × input_price) + (output_tokens ÷ 1,000,000 × output_price). At 2,000 input plus 500 output tokens on GPT-4o, cost is $0.01 per ticket. GPT-4o mini is $0.0006. Claude Haiku averages about $0.0037 per ticket at roughly 3,700 tokens per conversation.
The formula
Cost per support ticket is straightforward once you separate input and output tokens. Every ticket run has an input charge for the tokens you send and an output charge for the tokens the model returns. The exact number depends on the model used for that ticket, whether cached input pricing applies, and whether the request runs on standard or batch pricing.
Using 2,000 input tokens and 500 output tokens, here is the exact math across four common models. This makes it easier to compare raw API economics before vendor fees, containment assumptions, or staffing costs get layered in.
(2000/1,000,000 × $2.50) + (500/1,000,000 × $10.00) = $0.0050 + $0.0050 = $0.0100/ticket
(2000/1,000,000 × $0.15) + (500/1,000,000 × $0.60) = $0.0003 + $0.0003 = $0.0006/ticket
(2000/1,000,000 × $1.00) + (500/1,000,000 × $5.00) = $0.0020 + $0.0025 = $0.0045/ticket
(2000/1,000,000 × $3.00) + (500/1,000,000 × $15.00) = $0.0060 + $0.0075 = $0.0135/ticket
For many teams, the surprise is how low raw API cost looks compared with vendor contracts. That does not mean the API number is unimportant. It means routing, caching, and pricing structure can matter more than swapping one high-end model for another.
Benchmark table
| Model | Input $/MTok | Output $/MTok | Tokens/ticket | Cost/ticket | Cost/10K tickets |
|---|---|---|---|---|---|
| GPT-4o | $2.50 | $10.00 | 2,000 in / 500 out | $0.0100 | $100 |
| GPT-4o mini | $0.15 | $0.60 | 2,000 in / 500 out | $0.0006 | $6 |
| Claude Haiku 4.5 | $1.00 | $5.00 | ~3,700 avg (Anthropic) | $0.0037 | $37 |
| Claude Sonnet 4 | $3.00 | $15.00 | 2,000 in / 500 out | $0.0135 | $135 |
| GPT-4o (cached) | $1.25 | $10.00 | 2,000 in / 500 out | $0.0075 | $75 |
| GPT-4o mini (batch) | $0.075 | $0.30 | 2,000 in / 500 out | $0.0003 | $3 |
Source note: Anthropic's ~$37 per 10,000 tickets comes from published Claude API pricing documentation using Claude Haiku at about 3,700 tokens per conversation average.
Why token counts vary so much
Support ticket cost swings because support tickets are not all the same. Tier-1 tickets such as simple FAQ replies and status checks often stay in the 500 to 1,500 token range. Tier-2 tickets with multi-turn context usually land in the 1,500 to 4,000 range. Investigation tickets that pull account state, logs, and history can reach 4,000 to 15,000 tokens or more.
A mixed queue makes averages misleading. Consider 1,000 tickets made up of 700 tier-1 tickets at 1,000 total tokens and 300 investigation tickets at 8,000 input plus 1,500 output tokens on GPT-4o. The tier-1 calculation is (1000/1M × $2.50) + (300/1M × $10) = $0.0055 per ticket, or $3.85 for 700 tickets. The investigation calculation is (8000/1M × $2.50) + (1500/1M × $10) = $0.035 per ticket, or $10.50 for 300 tickets.
Total monthly API cost becomes $14.35 instead of the $10 figure you would get from a flat 2,000 input and 500 output assumption. On mixed queues, blanket token assumptions can understate actual costs by 40 percent or more because the expensive work is clustered in a smaller set of tickets.
Four ways to lower cost per ticket
| Method | Cost lever | Estimated reduction on applicable tickets |
|---|---|---|
| Model routing (mini for tier-1) | Model price difference (16.7×) | 50-70% on routed tier-1 volume |
| Prompt caching (system prompt) | Cached input price (50% off OpenAI, ~90% off Anthropic) | Up to 90% on cached tokens |
| Batch API for non-real-time | 50% discount on all tokens | 50% on batched calls |
| Token audit + trimming | Token count reduction | 20-40% on typical over-specified prompts |
Each method attacks a different cost lever. Routing changes model price, caching changes the price of shared input context, batching cuts the rate for work that can wait, and token trimming reduces the amount of billable text altogether. Teams that combine all four often reduce API cost per ticket by 60 to 80 percent on mixed support queues.
Cost per ticket vs cost per investigation
Cost per ticket is a useful overview metric, but it hides where money is really being spent. Simple deflections drag the average down. A queue with many easy tickets can look cheap even if the hard tickets remain expensive. That is why B2B technical support teams should also track cost per investigation.
Cost per investigation isolates the tickets that need real work: API spend on investigated tickets divided by investigations completed. A team with 70 percent deflection and 30 percent investigation might report a $0.002 average cost per ticket while still spending $0.025 per investigation. The second number is the one that shows whether routing, prompts, and workflow changes are working.
Benchmark your queue with your own ticket mix
Book a 30-minute demo or model your current volume, investigation rate, and vendor pricing in the calculator before you change tooling.
FAQ
What is the average LLM cost per support ticket?
It depends on model and token volume. At 2,000 input and 500 output tokens, GPT-4o costs $0.01 per ticket, GPT-4o mini costs $0.0006, and Claude Haiku costs $0.0045. Anthropic's published data puts Claude Haiku at about $37 per 10,000 tickets at roughly 3,700 tokens per conversation average.
What is the cheapest LLM for support automation?
GPT-4o mini at $0.15 and $0.60 per million tokens and Claude Haiku 4.5 at $1 and $5 per million tokens are the lowest published prices among capable models listed here. With OpenAI's Batch API discount, GPT-4o mini drops to $0.075 and $0.30 per million tokens, or about $3 per 10,000 tickets at a 2,000/500 token average.
Does prompt caching really cut costs by 90%?
Prompt caching lowers the cost of input tokens that hit the cache. OpenAI charges $1.25 per million cached GPT-4o input tokens, which is 50 percent off the standard $2.50 rate. Anthropic charges 0.1× base price, or about 90 percent off. Total savings depend on what share of your tokens are cacheable, often 40 to 70 percent for teams with shared system prompts.
How do you track cost per support ticket in practice?
Log input_tokens and output_tokens from each API response. Multiply by per-token price. Aggregate the result by ticket, then divide total API cost by ticket count. Most teams find API cost is 10 to 20 times lower than vendor per-resolution pricing, so vendor markup is usually where the real cost question sits.