AI & Services

A Small Experimental Model Beat Opus 4.8 on Three Benchmarks.

DeepSeek's experimental multimodal model V4-Flash-Vision-Exp, released August 21, narrowly beat Opus 4.8 on DeepSWE, Agents' Last Exam, and ZeroBench. It's available now with no price increase.

· Huuush

A Small Experimental Model Beat Opus 4.8 on Three Benchmarks

Key takeaways

  1. 1On August 21, DeepSeek officially released the experimental multimodal model V4-Flash-Vision-Exp.
  2. 2According to the official benchmark table, it narrowly beat Opus 4.8 on DeepSWE, Agents' Last Exam, and ZeroBench Pass@5.
  3. 3Adding image input carries the same pricing as the existing V4-Flash.
Table comparing benchmark scores for three models: DeepSeek V4-Flash-Vision-Exp, V4-Flash-0731, and Opus-4.8

Three places it beat Opus 4.8

DeepSeek announced via its official X account on August 21 that it had released the experimental multimodal model DeepSeek-V4-Flash-Vision-Exp on the DeepSeek API platform. The benchmark table published in the official API docs shows that the model scored 59.3 on DeepSWE, a coding and agent evaluation, narrowly ahead of Anthropic's Opus-4.8 (58.0). It also scored higher than Opus-4.8 on the multimodal evaluation Agents' Last Exam (27.3 vs. 25.7) and on ZeroBench Pass@5 (35.0 vs. 34.0).

It didn't lead across the board, though. In the same table, NL2Repo is 57.7, 12 points behind Opus-4.8 (69.7), and DSBench-Hard is 63.6, 8.1 points behind Opus-4.8 (71.7). With narrow leads on just three items, it's too early to call it "Opus-class." It's more accurate to say it has reached a level on par with large models on certain multimodal agent tasks.

Text performance is unchanged

DeepSeek says V4-Flash-Vision-Exp is on par with the existing V4-Flash-0731 on text abilities, including agents, reasoning, and world knowledge. That means this update adds the ability to read images and screens without cutting text performance. In fact, on the text-based items in the table (Terminal Bench 2.1, Cybergym, Toolathlon-Verified, and others), the score gap between the two models is mostly within 1 to 5 points.

If you already use V4-Flash for text-only workflows, there's no particular reason to switch now. The Vision version matters for agent tasks that also handle image-based input such as screenshots, tables, and charts.

It plugs in with no price increase

The new model can be called right away under the name deepseek-v4-flash-vision-exp through the Chat Completions, Messages, and Responses APIs, and images can be passed in three ways: base64, an external URL, or the Files API. DeepSeek says adding image input carries the same per-token price as the existing V4-Flash, with no separate price increase. DeepSeek Harness 0.1.1, released the same day, supports the new model by default, so you can try it with no extra setup.

Frequently asked questions

Is V4-Flash-Vision-Exp more expensive than the existing V4-Flash?
No. DeepSeek says the same pricing as the existing V4-Flash applies even with image input. There is no separate increase.
Is it really better than Opus 4.8?
Only on some benchmarks. It scored narrowly higher than Opus 4.8 on DeepSWE, Agents' Last Exam, and ZeroBench (Pass@5), but Opus 4.8 is still 8 to 12 points ahead on others such as NL2Repo and DSBench-Hard.
Can I use it right now?
Yes. You can call it as 'deepseek-v4-flash-vision-exp' through the Chat Completions, Messages, and Responses APIs, and images can be passed in via base64, an external URL, or the Files API. DeepSeek Harness 0.1.1, released the same day, supports it by default.
Does it affect existing text-only work?
DeepSeek says V4-Flash-Vision-Exp is on par with the existing V4-Flash on text abilities, including agents, reasoning, and world knowledge. There's no need to change text-only workflows.

Related articles