AI agents · OpenClaw · self-hosting · automation

Quick Answer

What Is DeepSeek V4 Flash Vision Exp? Cheap Vision API

Published:

The Short Answer

DeepSeek-V4-Flash-Vision-Exp is DeepSeek’s first vision-capable model in the V4 Flash line, released on August 21, 2026 under the API ID deepseek-v4-flash-vision-exp.

The pitch is narrow and specific: match V4 Flash on everything text-based — agents, reasoning, world knowledge — while adding image input at the same token pricing. That puts image understanding at $0.22/$0.66 per MTok off-peak, one of the cheapest vision endpoints available from a frontier-adjacent lab.

The catch is in the name. exp means experimental, and there are no open weights.

Key Facts

DeepSeek V4 Flash Vision Exp
ReleasedAugust 21, 2026
API IDdeepseek-v4-flash-vision-exp
StatusExperimental
Input typesText + image
Text parityMatches V4 Flash (agents, reasoning, knowledge)
Price (off-peak)$0.22 in / $0.66 out per MTok
Price (peak)$0.44 in / $1.32 out per MTok
Cache-hit input$0.007 off-peak / $0.014 peak
Image billingSame token pricing as text
Open weights❌ Not released

Last verified: August 29, 2026.

The Peak/Off-Peak Trap

This is the single most misquoted fact about DeepSeek pricing, and it applies here.

Since 16:00 UTC on August 16, 2026, DeepSeek V4 models use peak and off-peak rates, where off-peak is exactly half of peak. There is no flat price to quote.

WindowHours (UTC)Rate
Peak01:00-04:00 and 06:00-10:00$0.44 / $1.32
Off-peakall other 17 hours$0.22 / $0.66

For US and European working hours, off-peak is the realistic default. For Asia-Pacific business hours, much of your traffic lands in peak and your effective cost doubles. Batch jobs should be scheduled deliberately.

You will still find pages quoting $0.14 per million input tokens for V4 Flash. That was the pre-August-16 flat rate and it is dead. Any comparison built on it understates DeepSeek’s cost by roughly 57%.

What “Text Parity” Actually Buys You

DeepSeek’s framing is that the vision variant does not trade away text performance to gain image support. That matters more than it sounds.

The usual pattern with cheap vision models is a capability cliff: the model reads an image adequately but reasons worse, calls tools worse, and follows instructions worse than its text-only sibling. If a single model handles both modalities at the same standard, you can collapse a two-model pipeline into one.

Concretely, that removes an entire architectural layer from workflows like:

  • Document extraction — read a scanned invoice and reason about the values in one call.
  • Chart and dashboard analysis — interpret a figure and write the summary without handing off.
  • Screenshot-driven agents — a UI agent that sees the screen and decides the next action in the same context.
  • Mixed-corpus RAG — retrieve across PDFs containing both text and diagrams without a separate vision pass.

DeepSeek claims the model closes most of the multimodal gap to Anthropic’s Opus 4.8. That is a vendor-internal comparison and should be treated as directional, not as an independent benchmark result.

What “Experimental” Costs You

The exp suffix is a real constraint, not marketing.

Model IDs can change. Experimental endpoints are frequently renamed or folded into a GA release under a different identifier. Code pinning deepseek-v4-flash-vision-exp may need editing on short notice.

Behaviour can shift. Output formatting, refusal behaviour and tool-calling conventions are not frozen the way a GA endpoint’s are. Prompts tuned tightly against it may drift.

Availability is not guaranteed. There is no stability commitment behind an experimental endpoint.

No weights, no escape hatch. DeepSeek has published open weights for other V4 models, but not this one. If the endpoint changes or disappears, you cannot self-host your way out. That is the meaningful difference between this and something like GLM-5.3-Flash, which shipped MIT-licensed weights on day one.

The practical posture: prototype freely, deploy with a pinned fallback, and keep the prompt layer model-agnostic enough that swapping the vision backend is a config change.

Where It Sits in the DeepSeek V4 Line

As of late August 2026 the hosted V4 family exposes three distinct API IDs:

API IDModelStatusOff-peak price
deepseek-v4-proV4-Pro-0813GA$0.66 / $1.98
deepseek-v4-flashV4-Flash-0731Beta$0.22 / $0.66
deepseek-v4-flash-vision-expV4-Flash-Vision-ExpExperimental$0.22 / $0.66

The vision variant is priced identically to text-only Flash. There is no multimodal surcharge, which is unusual — most vendors either price vision models higher or bill image tokens at a different rate.

Who Should Use It

Good fit: high-volume document and screenshot pipelines where per-call cost dominates; prototypes that need image understanding without a separate vision model; teams already running DeepSeek V4 Flash for text who want to extend the same pipeline to images.

Poor fit: production systems needing endpoint stability guarantees; workloads requiring self-hosted weights for compliance or data residency; video understanding, which this model does not offer — GLM-5.3-Flash and Gemini remain the options there.

The honest framing: this is the cheapest credible way to add image understanding to an existing agent stack in August 2026, and it comes with exactly the reliability profile that price implies.

Sources