DeepSeek costs and claims: every figure from a sentence that names it

16 of the 629 figures in these field notes come from a sentence that names DeepSeek. Each one is quoted verbatim from the write-up it appeared in, with the day that write-up went out.

Some of these are about DeepSeek and some only measure against it — a competitor’s price quoted next to DeepSeek’s belongs here too, because that is the sentence someone searching for the comparison is looking for. The sentence tells you which is which.

This is not a DeepSeek price sheet. Nothing here is read off a vendor page. These are figures a working developer wrote down, and the date is the day the piece was published — not the day the price was in force. Model prices move fast: read every row as of its own date, and treat an old one as a lead rather than a quote.

The figures

90% — “The real power shows up in daily use: you can send complex logic to GLM, which performs at about 90% of GPT-level quality, while offloading routine tasks to cheaper models like DeepSeek V4 Flash.” (2026-09-30) →
70% — “Subtitle: How switching to DeepSeek and GLM saved me 70% and gave me back control” (2026-09-30) →
$1 per million tokens — “Even at the high end, models like GLM 5.2 and DeepSeek V4 Pro hover around $1 per million tokens.” (2026-09-30) →
$0.028 per million tokens — “One developer swapped their default backend for DeepSeek V4 Flash and saw API costs drop to just $0.028 per million tokens.” (2026-09-30) →
$0.028 per million input tokens — “For example, DeepSeek is known for high cost efficiency at $0.028 per million input tokens.” (2026-09-30) →
$0.028 per million tokens — “With DeepSeek V4 Flash priced at just $0.028 per million tokens, developers can assign lightweight tasks like boilerplate generation or syntax checking to this model, reserving higher-cost models for nuanced logic or debugging.” (2026-09-30) →
$0.028 per million tokens — “Take DeepSeek V4 Flash: it costs just $0.028 per million tokens.” (2026-09-30) →
90% — “GLM provides an experience benchmarked at 90% of GPT capabilities for daily tasks, while DeepSeek offers a highly competitive pay-as-you-go pricing structure.” (2026-09-25) →
80% — “A local proxy called OpenCodex redirects Codex requests to Chinese models such as DeepSeek and Zhipu GLM, which sidesteps the five-hour quota window and reportedly cuts cost by more than 80% — 10 yuan of DeepSeek credit lasting several days, with measured latency in the 100–200ms range and model switching that doesn’t break conversation context.” (2026-09-16) →
$0.028 per million tokens — “DeepSeek V4 Flash costs $0.028 per million tokens and is fast.” (2026-09-15) →
8.33% — “The row in my table said 8.33%, because the host my table recorded was StreamLake, not DeepSeek.” (2026-08-24) →
3.33% — “DeepSeek V4 Pro has seventeen, 3.33% to 20.00%.” (2026-08-24) →
3.33% — “DeepSeek publishes 0.022 per million for a cache hit against 0.66 cache miss — a ratio of 3.33%.” (2026-08-24) →
$0.0566 — “Gemini 3.7 Flash is the cheapest of the 40 at $0.0566 effective, and DeepSeek V4 Pro sits right next to it — a fact that turned out to be the loose thread in this whole piece.” (2026-08-24) →
$1 per million tokens — “Top-tier Chinese models such as GLM5.2 and DeepSeek V4 Pro sit near $1 per million tokens at inference gross margins of 10% to 20%.” (2026-08-19) →
$1 per million tokens — “DeepSeek V4 Pro and GLM5.2 charge roughly $1 per million tokens and reportedly run at a 10 to 20 percent margin.” (2026-08-05) →

All 629 figures, every kind · JSON · CSV

Where these 16 came from

Did this answer it? A star on the repository is the whole ask — the data is CC BY and does not require starring.

A DeepSeek figure that is not here yet? Say which metric, which unit, and where you read it — in one line. The form already knows it is about DeepSeek.

Or is one of the 16 above already out of date? Say which one — the form already knows it is about DeepSeek; you only have to say what the number is now.

All write-ups