Cost per task is the unit metric AI startups should track. It is the full cost of one finished customer job, not the price of one model call. Token price alone cannot set pricing or margins. A cheap model that retries often can cost more than a costly model that finishes cleanly.
Table of Contents
- Why token price misleads
- How large the model spread has become
- How to calculate it by workload
- How to hold the line on margin
Why token price misleads
A customer task rarely equals one model call. OpenAI advises budgeting each agent run separately because one task may call the model several times. That makes dollars per million tokens a weak guide to dollars per ticket, report, or workflow.
OpenAI and nOps guidance points to the same fix. Add inference plus tool and review expense for a workload, then divide by successful completions. Failures still count, because retries and fixes spend budget without earning revenue.
How large the model spread has become
Artificial Analysis found cost per benchmark task varies more than 13x among leading models, as reported in the Artificial Analysis Intelligence Index v5. In that test, GPT-6.1 Terra costs $0.96 per task and Claude Fable 5.1 costs $13.04. The middle tier shows the pricing power. Artificial Analysis found Claude Opus 5.5 scores 58 at $10.04 per task while GPT-6.1 Sol scores 55 at $1.27.
That is within four points at about one-eighth the task cost. Task tests show the same pattern, according to the OpenAI GPT-6.1 Sol announcement. OpenAI reports Sol beats Opus 5.5 on GDP.pdf questions at less than half the cost per task and on AutomationBench at roughly one-third. OpenAI lists Sol at $2 per million input tokens, $10 per million output, and $0.10 per million cached input.
How to calculate it by workload
Measure by workload, not by company average. Tickets, summaries, and agent workflows use different steps and succeed at different rates. A blended average hides the loss-making job.
Pull one number each week for each main job. Use these steps: This formula follows practical guidance from OpenAI and nOps. It forces retries, tool fees, and human fixes into the price.
- Total inference, tool calls, and review minutes for the workload.
- Divide by successful completions only.
- Set a per-task budget and step cap, then route simple jobs to cheaper models.
How to hold the line on margin
AI cost can move gross margin fast. nOps reports median SaaS gross margin around 77%, with several vertical providers disclosing 6-9 points of year-over-year compression after adding inference costs, in the nOps SaaS FinOps analysis. Set a hard cap before scaling volume.
OpenAI advises reserving the maximum possible next-request cost and stopping when the balance cannot cover it, as explained in the OpenAI per-run spending guide. Artificial Analysis notes its task prices are list-price calculations across input, caching, reasoning, and answer tokens in its own tests. Your discounts, caching, routing, and retry rates will differ, so block any launch that breaks its per-task budget.