DeepSeek-V4-Flash-Vision-Exp: Vision at Flash Price

DeepSeek shipped a vision SKU, not a new Flash. On August 21, 2026 it added deepseek-v4-flash-vision-exp to the official pricing table and changelog. Experimental. Same 1M / 384K window as V4-Flash and V4-Pro. Same Flash dollars, including the peak/off-peak split that started August 16.

The text-agent numbers they published sit next to Flash-0731: Terminal Bench 2.1 83.9 vs 82.7, DeepSWE 59.3 vs 54.4. That is the “on par” claim, and the deltas are small. The reason this is a post is the input type: images at Flash price, plus Chartography 64.3 and ZeroBench Pass@5 35.0, and a lab line that multimodal agents are close to Opus 4.8. They did not publish the Opus column.

At a Glance

• Release: August 21, 2026. Experimental.
• API / catalog ID: deepseek-v4-flash-vision-exp
• MODEL VERSION: DeepSeek-V4-Flash-Vision-Exp
• Modalities: text + image in, text out
• Context / max out: 1M / 384K
• Thinking: on by default; low / high / max
• Features: JSON, tools, Responses API, Anthropic API, Chat Prefix Completion (Beta). FIM not supported.
• Concurrency: 2500
• Off-peak: $0.22 miss / $0.007 hit / $0.66 out
• Peak (01:00-04:00 and 06:00-10:00 UTC): $0.44 / $0.014 / $1.32
• Images: billed as input tokens (about 384 tokens max per image after resize)

What's New

First official V4-Flash vision understanding API. Not a new text family. Experimental, accessed by setting the model string.

Same price as V4-Flash after the August 16 peak/off-peak hike. You are not paying a vision premium on the list.

Self-reported multimodal agent scores, and a “close to Opus-4.8” claim without an Opus column.

Benchmarks

All self-reported on the Aug 21, 2026 changelog. Not LLM Stats verified.

This is the on-par story. Terminal Bench +1.2, DeepSWE +4.9, AutomationBench +0.6. If you wanted a smarter text Flash, this is not that drop.

Chartography 64.3 and ZeroBench Pass@5 35.0 are the vision-adjacent numbers. ApexBench Pass@1 36.5 and Agents' Last Exam 27.3 sit in a footnote: text Flash ignores multimodal elements on those two, so the leap is not a clean paired score. DSBench-Hard 63.6 vs 59.6 is internal.

“Close to Opus-4.8” is a sentence, not a table. Do not invent Opus 4.8 scores to complete it.

Harness note from the changelog: Code Agent public benches used DeepSeek Harness minimal mode, max effort, topp=0.95, temperature=1.0.

Pricing

Same as V4-Flash. From the official pricing table, per 1M tokens:

Peak hours are 01:00-04:00 and 06:00-10:00 UTC. Off-peak is half of peak. Images convert to tokens by size and ride the input rate. detail=low downscales to 512×512 when you do not need fine pixels.

Context only: the August 16 hike took effect at 16:00 UTC that day. This SKU never sold at the old $0.14 / $0.28 Flash list.

How Vision Works

Three input paths: base64, public URL, or Files API file_id. Images in user messages only. Formats: JPEG, PNG, GIF, WebP. Max 600 images per request; 32 MiB per image (64 MiB via Files API). Other DeepSeek model ids return 400 on images. See DeepSeek's Vision guide for token usage and detail levels (low / high / original / auto).

When to Use It

Good fit

• Flash-priced screenshot, chart, or UI loops
• Codex or Responses clients that need image input
• Shops already on V4-Flash that do not want a Pro bill

Prefer text V4-Flash

When you have no images: same price, and FIM is available in non-thinking mode on text Flash (Vision-Exp does not support FIM).

Prefer V4-Pro

When you want the Pro agent stack and will pay roughly 3× list with 500 concurrency.

Do not treat experimental as GA. The id says Exp.

Migrating

• New id: deepseek-v4-flash-vision-exp. Text Flash does not accept images.
• Re-measure token-per-task. Images add input tokens (cap ~384 each).
• Thinking is on by default; effort low / high / max.
• Do not send images on deepseek-v4-flash or deepseek-v4-pro (400).
• Budget peak hours. Same clock as Flash.

Outlook

A vision SKU at Flash price is the news. The text benches do not move the Flash story. What will age this post: a non-exp GA id, a real Opus 4.8 comparison table, or an official weight drop.

Sources: changelog, pricing, vision guide.