Here is one thing that doesn’t happen everyday.
An AI company creates a model which has the ability to see and understand images, and does not ask you to pay any extra charge for it.
DeepSeek just did exactly that.
DeepSeek-V4-Flash-Vision-Exp was released on August 21, 2026. It is a multimodal model that adds visual understanding capabilities into their budget-friendly V4-Flash lineup.
And here's the kicker. It costs exactly the same as the text-only version.
That means you get powerful multimodal AI without paying a premium. The same $0.22 per million input tokens off-peak. The same $0.66 per million output tokens. No vision markup.
Let's break down what this actually means.
What Makes This Model Different
This isn't a completely new model. It's the same V4-Flash text model you already know, but with a vision encoder added to it.
Think of it like a pair of eyes on a smart assistant. The brain is the same. The intelligence is the same. But now it can actually see what you're showing it.
What stays the same:
1 million token context window
384,000 token max output
Same peak/off-peak pricing structure
Same text capabilities; agents, reasoning, world knowledge
What's new:
Image input support (JPEG, PNG, GIF, WebP)
Visual understanding for agent workflows
Multimodal API support
The text model and the vision model share everything except the ability to process images. The vision encoder sits in front of the same text brain.
How Well Does It Actually Perform?
DeepSeek released benchmark data showing how the vision model stacks up against both V4-Flash and Anthropic's Claude Opus 4.8.
On text-only tasks, V4-Flash-Vision-Exp matches the base model. No drop in quality. No trade-off.
On visual agent benchmarks, the improvement is significant.
Here are the numbers from the official announcement:
Benchmark | Vision Exp | V4-Flash Text | Opus 4.8 |
Terminal Bench 2.1 | 83.9 | 82.7 | 85.0 |
ApexBench (visual) | 36.5 | 26.2* | 39.4 |
Agents' Last Exam | 27.3 | 25.2* | 25.7 |
ZeroBench (visual) | 35.0 | - | 34.0 |
DSBench-Hard | 63.6 | 59.6 | 71.7 |
Chartography (visual) | 64.3 | - | 65.0 |
*The text-only V4-Flash ignores multimodal elements in these benchmarks.
The vision model nearly catches up to Claude Opus 4.8 on visual agent tasks. It actually beats Opus 4.8 on Agents' Last Exam and ZeroBench.
DeepSeek says the multimodal agent performance is "close to Opus-4.8". That's significant considering the price difference.
Image Processing and Billing
Here's where things get interesting.
Each image you send is tokenized and billed at the same rate as text input. But there's a catch—and it's actually good for you.
Token limits for images:
Each image costs up to 384 tokens maximum
This applies regardless of resolution
A 2000×2000 image costs the same as a 5000×5000 image
Image scaling:
Images are resized before processing
Final scaling approximates 800×800 pixels
This keeps token costs predictable and low
Practical cost:
Off-peak: ~$0.0085 cents per image
Peak: ~$0.017 cents per image
600 images in one request costs about $0.05-$0.10 total
The image processing is designed to be efficient. Large images are scaled down. Small images are scaled up. Everything lands at roughly 384 tokens per image.
What You Can Build With It
The vision model opens up entirely new use cases for the V4-Flash API.
UI Testing and Screenshot Analysis
Your agents can now analyze what's actually on screen. They can read error messages, interpret interfaces, and validate visual elements.
Chart and Diagram Extraction
Need to pull data from graphs? The model can read charts, understand diagrams, and extract information.
Document Screenshots
Upload a screenshot of a document and have the model process it. Useful for workflows that involve image-based documents.
Web Browsing Agents
Agents can now see the pages they're browsing. They can analyze visual layouts and read content from screenshots.
Error Message Capture
Users can share screenshots of errors. The model can read and diagnose issues from visual input.
Sample Use Cases from DeepSeek:
Generate a travel presentation with real photographic visuals
Recreate websites with specific design themes
Build interactive front-end demos from visual inspiration
API Access and Files API
Model Access:
Set model='deepseek-v4-flash-vision-exp' in your API calls.
Supported Formats:
Chat Completions
Messages
Responses API
Image Input Methods:
Base64 inline
External URLs
Files API (new)
New: Files API
DeepSeek also launched a free Files API alongside the vision model.
Upload an image once, reference it by file_id in multiple requests
64 MiB per file maximum
25 GiB total storage per user
10,000 files maximum per user
Expiration from 1 hour to 30 days
API itself is free to use
Why This Matters
DeepSeek's move is interesting for a few reasons.
No Vision Premium
Most companies charge extra for vision capabilities. DeepSeek didn't. The only cost is the image's token count.
Agent Focus
The model is clearly aimed at developers building autonomous agents. The vision capability helps agents interact with visual interfaces and understand graphical information.
Experimental Status
The model is marked "experimental" with no GA date yet. DeepSeek says they'll decide on permanent availability based on usage.
Weight Availability
The vision weights are not available for download. Only the text version (V4-Flash-0731) is on Hugging Face under MIT license.
When to Use Which Model
Choose V4-Flash-Vision-Exp when:
Your pipeline can ever receive an image
You're building UI-testing or screenshot-based agents
You need to read charts, diagrams, or screen captures
You want visual understanding without paying extra
Choose V4-Flash text-only when:
Your traffic is pure text and will never change
You need the confirmed stability of the GA version
You're working on coding completion or retrieval tasks
The vision model is strictly better at the same price for visual tasks. There is no text-quality penalty. The only reason to avoid it is if you need the stability of the non-experimental version.
FAQ Section
What is DeepSeek-V4-Flash-Vision-Exp?
It's an experimental multimodal model from DeepSeek that adds visual understanding capabilities to the V4-Flash text model. Its release date was August 21, 2026
How much does DeepSeek-V4-Flash-Vision-Exp cost?
It costs exactly the same as V4-Flash. Off-peak: $0.22 per million input tokens, $0.66 per million output tokens. Peak: $0.44 per million input, $1.32 per million output.
How are images billed?
Images are converted to tokens and billed at the same rate as input tokens. Each image costs up to 384 tokens maximum.
How many images can I send in one request?
Up to 600 images per request based on size.
What formats of images can be used?
JPEG, PNG, GIF, WebP.
How is image quality ensured?
The image is resized to an approximate size of 800×800 px. Thus, token costs are predictable.
What's the Files API?
A free API that lets you upload images once and reference them by file_id across multiple requests. 64 MiB per file, 25 GiB total storage.
Is this model available for download?
No. Only the text version V4-Flash-0731 is available on Hugging Face. The vision weights are API-only.
How does it compare to Claude?
The vision model is close to Claude Opus 4.8 on multimodal agent benchmarks, sometimes beating it. On text tasks, it matches V4-Flash.
Is the model permanent?
It's experimental. DeepSeek says they'll decide on making it a permanent offering based on usage.
