DeepSeek's new model can see images now, but what really matters isn't that it caught up with anyone. I scrolled past DeepSeek releasing a new model this morning, and my first reaction was: "Oh, another one that can see images." After reading through the official docs and benchmark tables, I changed my mind. The key point here isn't the "can see images" part. Models that can see images are a dime a dozen these days—GPT can, Claude can, and so can Gemini.

💡 What You Will Learn

DeepSeek's new model can see images now, but what really matters isn't that it caught up with anyone. I scrolled past DeepSeek releasing a new model this morning, and my first reaction was: "Oh, anoth

📜 Table of Contents

• A Cheap Model That Can "See Your Screen" (Vision Capabilities Are Live)
• 384 Tokens Per Image (Pricing Is DeepSeek's Real Signature Move)
• Won 3 Out of 11 Benchmarks (Let's Lay the Data Out)
• Don't Overlook Those Three Letters "Exp" (The Boundaries of an Experimental Version)
• My Take (The Visual Agent Price War Has Begun)

DeepSeek's New Model Can See Images, But What Really Matters Isn't That It Caught Up

Scrolling through my feed this morning and seeing DeepSeek drop a new model, my first thought was: "Oh, another one that can see images."

After reading the official docs and benchmark tables, I've changed my mind.

The key here isn't the "can see images" part. Models with vision are a dime a dozen now — GPT can see, Claude can see, Gemini can see. DeepSeek's real move this time: adding vision to their cheapest model, V4-Flash, and then dropping it directly into the multimodal Agent arena, going head-to-head with Anthropic's Claude Opus 4.8.

A Cheap Model That Can "See Your Screen" (Vision Capabilities Are Live)

Let's start with what this model can actually do.

DeepSeek-V4-Flash-Vision-Exp — long name, but break it down and it's three layers: V4-Flash is DeepSeek's cheapest flagship text model, Vision means it's got vision added, and Exp marks it as an experimental version. It can now read base64 images, external image URLs, and reference images via file_id through the new Files API.

Still sounds like technical jargon. Let me translate: from now on, when you write an Agent script and toss it a screenshot of a computer screen, it can figure out which button is where, what the error message says, which column of numbers in the table is off — then call the appropriate tools to take action. Or toss it a financial report chart, and it can read the differences between bars without you manually typing in the numbers.

That's what "having eyes" really means for Agents. Over the past year or so, I've played with quite a few Agent frameworks, and one of the biggest pain points is that they can only read text. You ask it to operate a webpage, and it has to guess what the page looks like from the HTML source code; the moment the page gets redesigned, or the content is canvas-rendered, it's completely lost. Once it can understand screenshots, the Agent is looking at the same screen a human operator sees.

384 Tokens Per Image (Pricing Is DeepSeek's Real Signature Move)

Vision models have existed before, but they were expensive.

DeepSeek's pricing this time: each image counts as a maximum of 384 tokens, billed at V4-Flash rates. What's V4-Flash going for? Third-party research firms tested it this month — processing a million words costs roughly $0.87, while Anthropic's comparable model runs about $50. That's nearly a 60x difference.

The Files API that launched alongside is also free. Upload an image once, get a file_id, and reference it directly in subsequent requests without re-uploading every time. That's especially critical for batch tasks — think customer service systems analyzing 10,000 user screenshots a day; you'd save a chunk just on bandwidth costs.

So you see, benchmark comparisons are just the surface. What actually makes enterprise developers seriously consider this model is the arithmetic between "slightly behind" and "60x cheaper."

Won 3 Out of 11 Benchmarks (Let's Lay the Data Out)

The official release includes 11 multimodal Agent benchmarks, and let me do the math for you.

V4-Flash-Vision-Exp beats Opus 4.8 on 3: DeepSWE by 1.3 points, Agents' Last Exam by 1.6 points, and ZeroBench by 1.0 points.

Roughly tied or slightly behind on Toolathlon-Verified (75.9 vs 76.2), Chartography (64.3 vs 65.0), and Terminal Bench 2.1 (83.9 vs 85.0).

The bigger gaps are on two: NL2Repo at 57.7 vs 69.7 — a 12-point difference — and DSBench-Hard at 63.6 vs 71.7, an 8-point gap. These two are the kind of hardcore tasks where you're handed a complete code repository and have to find and fix problems yourself.

I've always told my friends working on Agents: don't just look at the one or two benchmarks vendors cherry-pick for marketing. DeepSeek was relatively restrained this time — they published the full table, wins and losses clearly laid out. What you need to do is map it against your own use case: if you're running web operations, chart parsing, or terminal tasks, the gap is basically within a point or two; if your product's core is repository-level code refactoring, that 12-point gap might be the difference between shipping and not shipping.

Don't Overlook Those Three Letters "Exp" (The Boundaries of an Experimental Version)

A few details need to be clarified, or you'll get misled by clickbait headlines.

First, the "Exp" in the model name means experimental. DeepSeek itself labels it as an experimental version, which means the API could change, stability might fluctuate, and you should be cautious about putting it in production.

Second, it's benchmarked against Opus 4.8, which Anthropic released in May and is still in service. But Anthropic already released the newer Claude Opus 5 on July 24. Opus 5 isn't in DeepSeek's table, and nobody across the internet has done that comparison yet. Saying "close to Claude" isn't wrong, but the accurate statement is close to Claude Opus 4.8 — not catching up to Anthropic's latest flagship.

Third, there's a footnote in the official table that's pretty honest: the text-only V4-Flash was tested on multimodal benchmarks by "ignoring image content." Sounds like a no-brainer, but it shows that part of the benchmark gap comes from the old model simply not being able to see images. The new model's scores jump once it gains eyes — ApexBench goes from 26.2 to 36.5, Agents' Last Exam from 25.2 to 27.3. That improvement is real, but you need to know what baseline it's climbing from.

On the flip side, DeepSeek also revealed something else: across 7 text-only benchmarks, the vision version actually scored higher than the text-only version on 6 of them — Toolathlon up 5.6 points, DeepSWE up 4.9 points. Adding eyes didn't sacrifice the original capabilities, and that takes real engineering effort.

My Take (The Visual Agent Price War Has Begun)

Harness 0.1.1 launched the same day as the model, and I think that's worth more attention than the benchmarks.

Harness is DeepSeek's own Agent framework, positioned similarly to OpenAI's Codex SDK. The new model getting native support in the framework on day one means DeepSeek wants to hand you the whole package: cheap model, vision capabilities, and Agent orchestration framework — three things bundled together. Change one line in the model parameter, and a text-only Agent can suddenly read screenshots. The barrier to entry has been pushed down that low.

I wrote about the official V4-Flash release before — when it overtook its own Pro version on Agent benchmarks, I predicted DeepSeek would make "Flash-level pricing + Pro-level capability" their long-term strategy. Now that strategy has advanced another step: Flash pricing + multimodality + Agent framework, all three pieces in place.

For those of us who deal with Agents every day, this is good news. Nobody wants to pay $50 for a million words when $0.87 can get it done. Whether it actually gets it done is another question, but at least the choice is back in our hands.

In 2026, are you willing to pay 60x more for a 12-point benchmark lead? Let's talk in the comments.