DeepSeek V4 Flash Vision Exp: Opus‑4.8‑level Multimodal Performance for Under $0.20 per 1,000 Images

Model Launch and Core Claims

DeepSeek released the experimental multimodal model deepseek‑v4‑flash‑vision‑exp on its API platform, marking the company's first vision‑enabled model. The text capabilities remain on par with V4‑Flash, while the multimodal agent performance approaches that of Opus‑4.8, all priced at the Flash tier.

Benchmark Results

Community tests show that the vision‑enabled model retains the same text performance as V4‑Flash‑0731. On the Terminal Bench 2.1 it scores 83.9, and on DeepSWE it reaches 59.3, slightly higher than the pure‑text version. Compared with Opus‑4.8 across 11 evaluations, DeepSeek ties or exceeds Opus on three metrics (DeepSWE 59.3 vs 58.0, Agents' Last Exam 27.3 vs 25.7, ZeroBench 35.0 vs 34.0) and remains close on the others.

For multimodal tasks, ApexBench rises from 26.2 (text‑only) to 36.5, and Agents' Last Exam improves from 25.2 to 27.3. Independent developer tests place the model alongside top performers such as Kimi K3, GPT 5.6 Sol, and Opus‑4.8.

Cost Analysis

Images are token‑priced, with a maximum of 384 tokens per picture. The model automatically rescales large inputs to an 800×800‑pixel budget and charges at the V4‑Flash text rate. Peak pricing is $0.44 per million tokens (idle $0.22), translating to roughly $0.00017 per image. Processing 1,000 images therefore costs under $0.20.

High‑resolution screenshots cost about $0.0002 each, compared with $0.0156 for Opus, a cost advantage of roughly 1/80.

Files API and Practical Tips

The newly added Files API is free and returns a file_id for each uploaded image, enabling cross‑request and cross‑session reuse without re‑uploading. Limits include a 64 MB maximum file size, 25 GB total storage, up to 10 k files, and a retention period of 1 hour to 30 days.

For production use, avoid embedding images as base64 to save bandwidth and stay within the 48 MB request‑body limit; instead, use the Files API.

Suitable Use Cases and Limitations

Community feedback suggests the model excels at screenshot‑driven visual agents, UI automation, document extraction, chart Q&A, batch OCR, and mixed long‑form image‑text office automation. It struggles with pixel‑level fine‑grained recognition and small‑font OCR, which still require dedicated OCR models.

Requests can contain up to 600 images, each up to 8 192 px (reduced to 4 096 px when more than 15 images are sent). Supported formats are JPEG, PNG, GIF, and WebP.

SDK Update

DeepSeek Harness v0.1.1 now supports the new V4‑Flash‑Vision‑Exp model and native image request configuration.

Reference links: https://x.com/deepseek_ai/status/2090730032574631962 and https://x.com/tianyi/status/2090730841509171466.