DeepSeek V4 Flash vs Pro: A Bulk Text Cost Receipt

01 / Start With 10,000 Documents, Not a Benchmark

For bulk text work, compare the receipt before debating model rankings. Suppose you need to extract fields, assign categories, and produce a short summary for 10,000 documents. Budget 2,000 input tokens and 500 billed output tokens per document: 20 million input tokens and 5 million output tokens altogether.

Those are planning assumptions, not measured performance. They include instructions in the input allowance and assume every document needs one request. Your pilot should replace both averages before you commit the full queue.

The useful question is not whether DeepSeek V4 Pro sounds more capable than DeepSeek V4 Flash. It is whether any improvement on your documents justifies the additional cost per accepted result.

02 / The Rate Card Behind This Receipt

DeepSeek's official pricing checked October 25, 2026 lists off-peak rates per million tokens of $0.15 uncached input and $0.60 output for Flash, versus $0.66 and $1.98 for Pro. These are direct-provider reference rates, not verified CheapRelay prices. ([api-docs.deepseek.com](https://api-docs.deepseek.com/quick_start/pricing/))

There is also a naming caveat: the official Flash alias now serves an updated underlying model. Confirm CheapRelay's actual model mapping before treating matching labels as identical deployments. ([api-docs.deepseek.com](https://api-docs.deepseek.com/quick_start/pricing/))

For this comparison, assume no cache hits, no separate cache-write charge, no retries, and identical token consumption. This isolates the price difference; it does not establish equal output quality.

03 / Receipt One: Run Every Document on Flash

• ■ Uncached input: 20 million × $0.15 = $3.00.
• ■ Billed output: 5 million × $0.60 = $3.00.
• ■ Total: $6.00, or $0.60 per 1,000 documents. ([api-docs.deepseek.com](https://api-docs.deepseek.com/quick_start/pricing/))

Use this as a candidate baseline for jobs with explicit acceptance rules: required fields present, allowed category selected, valid structure, and no unsupported claims. Do not assume the lower bill means the output is good enough. Test it against a small, manually reviewed set containing both ordinary documents and awkward edge cases.

04 / Receipt Two: Run Every Document on Pro

• ■ Uncached input: 20 million × $0.66 = $13.20.
• ■ Billed output: 5 million × $1.98 = $9.90.
• ■ Total: $23.10, or $2.31 per 1,000 documents.
• ■ Additional spend over Flash: $17.10. ([api-docs.deepseek.com](https://api-docs.deepseek.com/quick_start/pricing/))

That additional spend might be worthwhile if your own evaluation shows fewer consequential errors or less editing. It might also buy nothing useful for straightforward labeling. Record acceptance rates separately from API success rates: a completed response is not automatically an acceptable result.

05 / Receipt Three: Flash First, Pro for Exceptions

Assume a validation process sends 10% of documents to Pro after their initial Flash pass. That means paying for all 10,000 Flash requests plus 1,000 additional Pro requests, not replacing the original calls.

• ■ Flash pass: $6.00.
• ■ Pro escalation: 10% × $23.10 = $2.31.
• ■ Combined total: $8.31.
• ■ Saving versus Pro throughout: $14.79, approximately 64%. ([api-docs.deepseek.com](https://api-docs.deepseek.com/quick_start/pricing/))

This is a routing scenario, not a claimed rejection rate. Escalation prompts must fit the original token allowance for that receipt to hold. If you append failed responses or extra instructions, count those tokens too.

06 / Reconcile the CheapRelay Bill Before Scaling

CheapRelay offers pay-as-you-go access to both models through one key, with no monthly fee or subscription. Its billing counts actual tokens, including cache reads and writes. Substitute its displayed rates into this receipt rather than assuming the direct-provider totals apply.

Keep separate ledger entries for uncached input, cache reads, cache writes, billed output, and repeated requests. Track accepted documents alongside total spend. For operational budgeting, add your own review costs separately.

Sign up for CheapRelay, run a small labeled pilot on both models, and calculate cost per accepted document. Choose Flash throughout, Pro throughout, or selective escalation from that evidence—not from the model names.