An automated process that reads a 500,000-row spreadsheet every working day does not need a frontier model on every row. OfficeMaker can interrogate the workbook deterministically and pass only the relevant records to a cheaper model. On simplified input-token assumptions, that redesign cuts routine LLM input cost from about £4,060 a month to about £0.41, with GPT-6 Astra reserved for genuine exceptions.
The operating rule is simple: filter the data first, pick the cheapest model that can do the remaining work, and escalate only what is actually hard.
Why does the naïve path cost thousands a month?
Suppose an automated process needs to analyse a 500,000-row spreadsheet each working day. Assuming approximately 50 tokens per row, processing the whole dataset is about 500,000 × 50 = 25 million input tokens.
If the whole dataset were processed using GPT-6 Astra, currently priced at $10 per million input tokens, the theoretical fresh-input cost would be 25M × $10 = $250 ≈ £185 per run. Run every working day: £185 × 22 days ≈ £4,060 per month.
That is input processing alone, before output tokens, additional agent reasoning, tool calls or retries. Astra output tokens currently cost $50 per million. A 25 million token prompt would also cross Astra’s long-context pricing band, so this naïve figure is a lower bound.
What happens if you filter the spreadsheet first?
Redesign the same process properly. [OfficeMaker](https://officemaker.ai) first interrogates the spreadsheet deterministically and reduces the 500,000 rows to the 2,500 records actually relevant to the task. That step is not generative: the service queries a cached model of the workbook and returns compact JSON. The language model never sees the other 497,500 rows.
That represents 2,500 × 50 = 125,000 input tokens. If those records are then processed using GPT-5.6 Luna, currently priced at $0.20 per million input tokens: 0.125M × $0.20 = $0.025 ≈ £0.018 per run — less than 2 pence of Luna input tokens. Across 22 working days: about £0.41 per month.
Which model should do the routine work?
Luna is the default for the daily pass on the reduced set. The workflow can then escalate only genuinely difficult cases — perhaps 100–200 exceptions — to Astra for higher-quality reasoning. Model selection follows the data, not the other way around.
- Naïve approach: 500,000 rows → Astra → ~£4,060/month
- Engineered approach: 500,000 rows → OfficeMaker → 2,500 relevant rows → Luna → ~£0.41/month for routine input processing, with only exceptional cases escalated to Astra
That is roughly a 10,000-fold reduction in the routine LLM input cost in this illustrative example. The same discipline applies when generating Office files: keep structure on the server instead of looping screenshots through the model. See OfficeMaker vs Python: Token Efficiency and the product write-up [Query Excel Before You Send It to the LLM](https://officemaker.ai/blog/query-excel-before-the-llm).
How should a production workflow be designed?
- Convert the workbook once so the query layer can scan sheets without loading them into the chat.
- Ask deterministic questions: which rows match, which keys join, which aggregates matter.
- Send only that compact result to GPT-5.6 Luna for routine analysis.
- Route genuine exceptions to GPT-6 Astra.
- Apply updates with keyed edits and human approval, not a full-file rewrite.
AI Build designs this two-layer pattern for UK teams: a cheap, governed daily path, and a frontier model only where judgement actually changes the outcome. If you are planning a ChatGPT Business rollout that will touch operational spreadsheets, start from the data contract — not from dumping the file into the most expensive model.