Measure AI Cost per Completed Task

A model invoice does not tell you what one correctly completed request cost. Connect spending to the process, the attempts it took and the time spent reviewing the result.

Lower token prices do not guarantee a cheaper workflow. OpenAI announced on 30 July 2026 an 80% price reduction for GPT-5.6 Luna and a 20% reduction for Terra. Those are rate changes, not equivalent savings on every business task.

Count the calls as well as the tokens

A chatbot may answer in one call. An invoice workflow may read the document, propose a match and draft a reply separately. Every step and retry adds consumption.

For an illustrative calculation, use the Terra rates in that announcement, $2 per million input tokens and $12 per million output tokens. With 100,000 input and 10,000 output tokens, 0.1 × 2 + 0.01 × 12 = $0.32. Five such calls cost $1.60. Both the volume and call count are chosen assumptions. The example excludes tools and human work. It is not a current rate card.

Compare correctly completed tasks

Tag calls with a process identifier. Divide the trial's total cost by the tasks that met the agreed quality standard. Include unsuccessful attempts, review and corrections. Report unfinished tasks as well.

Output token charges cover only part of the comparison. Reasoning and tools may also be billed. Check the provider's terms.

Model settings belong in the measurement

Anthropic's Opus 5 announcement compares performance and cost across effort levels. These are vendor evaluations on selected tasks. Compare settings on your own process rather than choosing by model name alone.

Record the version, effort, input size and cache use. Check changes against the same evaluation set.

Put controls around the next pilot

Keep models and settings behind a shared interface so replacing one does not require hunting through the application. A configuration change still needs verification of behaviour.

Cap calls and spending per run. A monthly alert may not stop a repeating error soon enough. With caching, check billing and which inputs actually repeat.

Try a bounded task on the cheaper model first. A multi-model strategy helps if corrections do not consume the saving. Our AI development cost comparison covers the relationship to developer time.

Design measurement alongside the workflow when planning AI automation. Contact us if you need to establish the cost of an accepted result.

Sources

• Anthropic, Introducing Claude Opus 5
• OpenAI, Advancing the price-performance frontier with GPT-5.6
• VentureBeat, AI price wars, OpenAI cuts GPT-5.6 Luna prices by 80 percent
• CNBC, OpenAI cuts prices for two of its GPT-5.6 AI models
• FinOps Foundation, State of FinOps 2026
• Analytics Vidhya, July 2026 AI releases timeline
• DataCamp, Grok 4.5 features, benchmarks and pricing
• Axis Intelligence, AI inference cost statistics 2026
• OpenRouter, Kimi K3 model card