Everyday workflows

This lesson is the practical core of the course. Each workflow lists the steps, the model to use, and a cost estimate on the API using catalogue prices at the time of writing. On the free app, every one of these costs $0; the API numbers are there so you know what automation would add, and how little it usually is.

1. Writing and editing

Draft in your own words first, then paste and ask for one specific pass: tighten, restructure, adjust tone, or check for missing information. One pass per message. For a 1,500-word article that is roughly 2,000 input tokens and 2,000 output tokens per pass. Three passes on V3.2 ≈ 6K in + 6K out ≈ less than a cent. Use the chat model; the reasoner adds nothing to prose.

2. Research and summarising

Ten 8,000-token documents summarised individually on V4 Flash 0731 is 80K input and about 5K output ≈ 80 × $0.000065 + 5 × $0.00018 ≈ under a cent. Combining the ten summaries (about 6K tokens) into a brief on V4 Pro 0813 is 6K in and 1.5K out ≈ under a cent too. The whole research pipeline costs less than the coffee you drink while it runs. Turn Search on in the app only for the collection step.

3. Coding

DeepSeek's sweet spot. Use the chat model for writing functions, explaining code, generating tests and refactoring. Use the reasoner for debugging that the chat model gets wrong twice, and for algorithm design. Paste the smallest unit that reproduces the problem — a function and a failing test, not the repository.

A working day of IDE-assisted coding might be 150 requests at 4,000 input and 600 output tokens: 600K in and 90K out. On V3.2 that is 0.6 × $0.27 + 0.09 × $0.40 ≈ $0.20 a day, or about $4 a month of workdays. If a tenth of those requests escalate to R1 0528, add roughly 60K in and 30K out of reasoner traffic ≈ $0.10 a day. Compare that with what you pay for a coding subscription elsewhere: DeepSeek vs Claude for coding.

4. Data and spreadsheets

Paste a sample of rows plus the column names, not the whole sheet, and ask for the formula, the SQL, or the Python that does the transformation. Then run it yourself. Asking the model to process 50,000 rows directly is slow, error-prone and expensive; asking it to write the ten-line script that processes them is fast and nearly free. For a genuinely tricky query, escalate to the reasoner with the schema and one example of the desired output.

5. Planning and decisions

Give the model the constraints — budget, deadline, people, what has already been tried — and ask for options with trade-offs rather than a single answer. Then ask it to argue against its own top recommendation. This is one of the few everyday tasks where the reasoner is worth switching on, because weighing several constraints at once is exactly what it does better than the chat model.