How much does DeepSeek's vision model cost per image?

Images sent to deepseek-v4-flash-vision-exp are converted to tokens and billed as input, with a hard upper bound of 384 tokens per image. At the off-peak input rate of $0.22 per 1M tokens, a maximum-size image costs $0.0000845; at the peak rate of $0.44, $0.000169. A 2000×2000 and a 5000×5000 image cost the same because both are resized to roughly 800×800-equivalent pixels before inference.

What are DeepSeek's peak and off-peak hours?

Peak hours are 01:00-04:00 and 06:00-10:00 UTC, which is 09:00-12:00 and 14:00-18:00 Beijing time. All other hours are off-peak, and off-peak rates are exactly half of peak rates. The same windows apply to deepseek-v4-flash and deepseek-v4-flash-vision-exp. Update Aug 23, 2026: weekends (Saturday and Sunday) no longer have peak windows — all weekend hours bill at off-peak rates.

Is vision-exp more expensive than deepseek-v4-flash?

No. deepseek-v4-flash-vision-exp is priced identically to deepseek-v4-flash: input cache-hit $0.007 off-peak / $0.014 peak, input cache-miss $0.22 / $0.44, output $0.66 / $1.32 per 1M tokens. The only cost difference is that vision requests also carry image tokens, capped at 384 per image.

How do I reduce vision API costs on DeepSeek?

Four levers: run image-heavy batches in off-peak hours (half price); crop screenshots to the relevant region instead of sending full screens; put repeated images (logos, templates, UI chrome) early in the prompt so they hit the $0.007/1M cache-hit rate; and reuse files via the Files API file_id instead of re-uploading identical base64 payloads.

Guides / DeepSeek Vision API Pricing

Guide · verified 2026-08-21 against api-docs.deepseek.com

DeepSeek Vision API Pricing (2026): What an Image Actually Costs

DeepSeek's vision model deepseek-v4-flash-vision-exp (launched August 21, 2026) is priced identically to V4-Flash — and every image is billed at a hard maximum of 384 tokens. This page has the full rate table, the image tokenization rule, worked cost examples, and how vision-exp stacks up against Western vision APIs. Every number below was read from the official pricing and vision docs on Aug 21, 2026.

Direct answer

deepseek-v4-flash-vision-exp costs $0.22 per 1M input tokens off-peak and $0.44 peak, with images capped at 384 tokens each — so a maximum-size image costs at most $0.0000845 off-peak, and 1,000 screenshots cost about $0.08. Output is $0.66 off-peak / $1.32 peak per 1M tokens; cached input hits $0.007 / $0.014. Peak hours: UTC 01:00–04:00 & 06:00–10:00 (Beijing 09:00–12:00 & 14:00–18:00).

Full rate table (USD per 1M tokens)

Rate	Off-peak	Peak	RMB equivalent
Input — cache hit	$0.007	$0.014	≈ ¥0.05 / ¥0.10
Input — cache miss	$0.22	$0.44	≈ ¥1.5 / ¥3.0
Output	$0.66	$1.32	≈ ¥4.5 / ¥9.0
Concurrency limit	2,500 (same as V4-Flash; V4-Pro is 500)

Source: official pricing page, read 2026-08-21. Off-peak is exactly half of peak in every tier.

The image rule: how 384 tokens is computed

Resize before inference. Images below roughly 384×384 pixels are scaled up (aspect ratio preserved); larger images are scaled down to roughly 800×800-equivalent pixel area.
Hard cap at 384 tokens. Because of the resize, there is an upper bound of 384 tokens per image — a 2000×2000 screenshot and a 5000×5000 photo cost exactly the same.
Each image bills independently. In a multi-image request, every image gets its own ≤384-token charge; you can send up to 600 images per request.
Formats and sizes. JPEG / PNG / GIF / WebP; 32 MiB per image inline or by URL; 64 MiB via Files API file_id; max 8,192 px per side (4,096 px with 15+ images).

Worked examples: what your workload costs

Scenario (max-size images, cache miss)	Tokens	Off-peak	Peak
1 screenshot	≤384	$0.0000845	$0.000169
1,000 screenshots	≤384K	$0.0845	$0.169
10,000 screenshots	≤3.84M	$0.845	$1.69
1,000 repeated UI images (cache hit)	≤384K	$0.0027	$0.0054

Text tokens around the images (prompts, tool output, model replies) bill at the normal rates on the same request — the rows above isolate the image component.

vs Western vision APIs (2026 list prices)

Model	Input $/1M	vs vision-exp	Image billing
deepseek-v4-flash-vision-exp	$0.22 / $0.44	baseline	≤384 tokens/image, hard cap
Gemini 3.7 Flash (intro)	$0.75	1.7–3.4×	resolution-dependent, no published cap
Grok 4.6	$2.00	4.5–9.1×	resolution-dependent
Claude Sonnet 4.5	$3.00	6.8–13.6×	typical screenshot ≈1,000+ tokens → ≈$0.003+/image (estimate)

Competitor list prices from public pricing pages; their per-image token counts vary by resolution (estimates marked). Structural point: vision-exp pairs the cheapest per-token price in its class with the only hard per-image token cap.

Four ways to cut your vision bill

Shift batch jobs off-peak. Exactly half price in UTC 04:00–01:00 windows; for a screenshot-processing pipeline this is a 50% cut for zero code changes.
Crop, don't dump. A cropped UI region still bills ≤384 tokens but produces better answers than a full screen — cost floor is the same, quality goes up.
Put repeated images early. Cache-hit input is ~31× cheaper than cache-miss ($0.007 vs $0.22). Identical logos/templates/UI chrome at the start of the prompt hit the cache.
Reuse via Files API. Upload once, reference file_id — no re-uploaded base64 bloat in every request. See our Files API guide.

FAQ

How much is one image on DeepSeek vision?

At most 384 tokens → $0.0000845 off-peak, $0.000169 peak. Size above ~800×800-equivalent pixels doesn't increase the cost.

When are peak hours?

UTC 01:00–04:00 and 06:00–10:00 — Beijing 09:00–12:00 and 14:00–18:00. Everything else is half price.

Is vision-exp pricier than plain v4-flash?

No — identical rates. You only pay extra for the image tokens themselves, capped at 384 each.

Does the price change since it's "experimental"?

It could — experimental models can see pricing or terms changes on short notice. These figures are the official ones as of Aug 21, 2026; re-check the pricing page before locking budgets.

Launch brief

Vision-Exp Official Launch

Timeline, quickstart, model routing

Guide

DeepSeek Files API

Uploads, limits, file_id examples

Comparison

Chinese Vision Models 2026

Vision-Exp vs Qwen-VL vs Kimi

Guide

DeepSeek API guides

All DeepSeek integration notes