LLM API · DeepSeek
DeepSeek V4 Flash pricing
Input
$0.440 /Mtok
Output
$1.32 /Mtok
Cached input
$0.014 /Mtok
Context window
1,000k tokens
Max output
384k tokens
Tier
budget
Ranked 12 of 29 by price among 29 LLM APIs in the index — the cheapest is Mistral AI Ministral 3 (3B) at $0.100. Figures are USD list prices excluding tax.
Standard (peak) rate shown. Off-peak is 50% lower and applies outside 01:00-04:00 and 06:00-10:00 UTC, Mon-Fri.
Off-peak rate: input $0.22, output $0.66, cached input $0.007 per million tokens.
Price history
Every weekly reading since this item entered the index on 15 August 2026. Dots mark readings where the figure changed.
| Date | Figure | Reading | Change |
|---|---|---|---|
| 2026-08-15 | Cached input $/Mtok | $0.014 | — |
| 2026-08-15 | Input $/Mtok | $0.440 | — |
| 2026-08-15 | Output $/Mtok | $1.32 | — |
| 2026-09-05 | Cached input $/Mtok | $0.014 | — |
| 2026-09-05 | Input $/Mtok | $0.440 | — |
| 2026-09-05 | Output $/Mtok | $1.32 | — |
Machine-readable: history.json · Every movement also appears in the change log.
What it costs in practice
Computed from the current figures above. Change the assumptions in the LLM API cost calculator.
| Workload | Cost |
|---|---|
| One million input tokens | $0.440 |
| One million output tokens | $1.32 |
| Blended $/Mtok at 4:1 input:output | $0.616 |
| 1,000 chatbot replies (1,500 in / 300 out) | $1.06 |
| Summarising 1,000 ten-page documents (5,000 in / 400 out) | $2.73 |
| A month of 10,000 requests a day (2,000 in / 500 out) | $462.00 |
| 1,000 replies with a 1,200-token cached prefix (300 fresh in / 300 out) | $0.545 |
Compared with other models in the same tier
| Model | blended $/Mtok (4:1) | vs this |
|---|---|---|
| Mistral AI Ministral 3 (3B) | $0.100 | -84% |
| Groq GPT-OSS 20B | $0.120 | -81% |
| Google Gemini 2.5 Flash-Lite | $0.160 | -74% |
| Qwen2.5 7B Instruct Turbo via Together AI | $0.300 | -51% |
| Cohere Command-light | $0.360 | -42% |
How to read this page
The figures are the list prices published on the vendor's own pricing page, re-read every week by an automated check that never invents a number: if the page cannot be read, the previous reading is kept and the item is flagged after three missed weeks. Negotiated rates, volume tiers, free credits and regional pricing are not modelled. The methodology describes the process and its limits; the price index shows every llm api figure side by side.