Published: 22 August 2026
DeepSeek released DeepSeek-V4-Flash-Vision-Exp on August 21, 2026, an experimental multimodal version of its V4-Flash model that adds image and screenshot understanding. On DeepSeek’s own agent benchmarks it beats Anthropic’s Claude Opus 4.8 on 3 of 11 tests, including Agents Last Exam, 27.3 versus 25.7, while charging the same price as text-only V4-Flash.
What is DeepSeek-V4-Flash-Vision-Exp and what changed?
V4-Flash-Vision-Exp is a multimodal variant of DeepSeek’s existing V4-Flash model, first released July 31, 2026. The new checkpoint keeps V4-Flash’s text reasoning, agent tool use and 1 million token context window unchanged, and adds the ability to read images and screenshots directly. DeepSeek says the model matches V4-Flash on every text benchmark, so the vision addition costs nothing in text quality.
The model ships alongside DeepSeek Harness 0.1.1, an agent framework update with built-in support for the new checkpoint, and a free Files API that lets developers upload an image once and reuse it across requests by file ID instead of re-uploading. Both are aimed at agentic workflows that combine screen reading with tool calls.
How does it perform against Claude Opus 4.8?
DeepSeek published results on five multimodal and agentic benchmarks.
- V4-Flash-Vision-Exp scores 27.3 on Agents Last Exam against Opus 4.8’s 25.7,
- 35.0 on ZeroBench against Opus 4.8’s 34.0,
- On ApexBench Pass@1 it trails at 36.5 versus 39.4,
- On Terminal-Bench 2.1 it scores 83.9 against 85.0.
- The gap widens on repository level coding. NL2Repo puts V4-Flash-Vision-Exp at 57.7 against Opus 4.8’s 69.7, a 12 point deficit DeepSeek does not try to explain away.
Across the disclosed benchmarks the model wins on 3 of 11, so approaching Opus 4.8 is closer to the truth than beating it outright. DeepSeek ran the tests itself, in Harness Minimal Mode, with no independent verification yet.
What does it cost and how do you access it?
V4-Flash-Vision-Exp is live now on DeepSeek’s API under the identifier deepseek-v4-flash-vision-exp, and on OpenRouter. Pricing matches V4-Flash exactly: $0.44 per million input tokens and $1.32 per million output tokens at peak hours, half that off peak. Images are billed as input tokens, up to 384 tokens per image.
Unlike DeepSeek’s usual practice of publishing open weights on Hugging Face for V4-Flash and V4-Pro, the company has not released downloadable weights for this experimental checkpoint. For now it is API only, and DeepSeek has not said whether or when that will change.
What this means
Worth testing now for teams already on DeepSeek’s API who need an agent to read a screenshot or a chart without switching providers or paying Opus 4.8 prices. The 3 of 11 benchmark split is honest enough to trust directionally: this is a capable, cheap vision add-on, not a frontier multimodal model.
The NL2Repo gap is the real signal. For coding agents that need to reason across a full repository from screenshots, Opus 4.8 and GPT-5.6 Sol remain ahead. Teams building lightweight screen reading agents on a budget should test this model; teams doing serious visual coding work should wait for the gap to close.
For more information visit the official DeepSeek’s announcement of V4-Flash-Vision-Exp.
