DeepSeek quietly shipped an experimental vision model on August 21 that does something most of the AI industry has been working toward: it adds image understanding to a cheap, fast text model without gutting the text performance, and the result lands within striking distance of a frontier model on agent benchmarks.
The model, deepseek-v4-flash-vision-exp, takes the text-only V4-Flash and adds the ability to read images, screenshots, and charts. DeepSeek says it “matches V4-Flash on text capabilities, including agents, reasoning, and world knowledge” while bringing multimodal agent performance “close to Opus-4.8” — Anthropic’s current supported model from May. On three of eleven benchmarks DeepSeek published, the new model actually beats Opus-4.8. On the rest, it trails by margins ranging from one to twelve points.
The price gap is where this gets interesting. V4-Flash costs about $0.22 per million input tokens at peak rates. Anthropic’s Opus models run roughly $50 per million. That is not a rounding error. It is a 200x difference, and for developers building agent pipelines that process thousands of images, it is the difference between a prototype and a product.
What the numbers actually say
DeepSeek published eleven benchmark results comparing the new vision model against V4-Flash-0731 (the text-only baseline) and Opus-4.8. The split tells two stories.
On text-based agent benchmarks, the vision model holds its own. It actually improves on the text-only V4-Flash on six of seven text benchmarks — Toolathlon-Verified jumps 5.6 points, DeepSWE by 4.9. The one exception is Cybergym, a security benchmark where adding vision cost 1.4 points. The headline claim of “matching” text performance is, if anything, undersold.
On multimodal agent benchmarks, the leap is real but comes with an honest footnote. ApexBench jumps from 26.2 to 36.5. Agents’ Last Exam goes from 25.2 to 27.3 — edging past Opus-4.8’s 25.7. ZeroBench Pass@5 hits 35.0 against Opus-4.8’s 34.0. But DeepSeek’s own table notes that the text-only V4-Flash “ignores multimodal elements” in two of those tests. So part of the gap is what happens when you give a blind model an eye test. DeepSeek disclosed this in the table itself — more transparency than most labs offer.
The result that should give pause: NL2Repo, a repository-scale coding benchmark, sits at 57.7 against Opus-4.8’s 69.7. That is a 12-point gap on exactly the kind of task enterprises are buying agents to do.
The Opus question
Here is the catch, and DeepSeek knows it. Opus-4.8 is not Anthropic’s newest model. Claude Opus 5 launched on July 24, and nobody has published a head-to-head between V4-Flash-Vision-Exp and Opus 5. DeepSeek picked a current, supported competitor — Opus-4.8 is listed as Active on Anthropic’s deprecation page with no retirement before May 2027 — but it is not the frontier.
The Next Web’s coverage puts it plainly: “The test that would settle it has a name and no results. Somebody needs to run V4-Flash-Vision-Exp against Opus 5 on the same harness.” Until that happens, the honest summary is narrower than the headlines. An experimental Chinese multimodal model sits within a few points of a supported American model on most of a benchmark set, wins three of eleven, costs a fraction as much, and trails by twelve points on the hardest task in it.
All benchmarks come from DeepSeek’s own harness in minimal mode, with temperature 1.0 and top_p 0.95. Vendor benchmarks are standard practice. They are still the vendor’s numbers.
Why the API details matter commercially
DeepSeek shipped this with three API compatibility layers — OpenAI Chat Completions, Anthropic Messages, and OpenAI Responses. That is not a technical footnote. It means teams already integrated with Claude or GPT can swap in V4-Flash-Vision-Exp without rewriting their agent pipeline. A model that speaks three dialects removes the adapter layer that bloats production stacks.
There is no vision surcharge. Images bill at the same per-million-token rate as text, capped at 384 tokens per image regardless of resolution. A 2000×2000 screenshot and a 5000×5000 photo cost the same. For agent workflows that process dozens of screenshots per task, that pricing model is aggressive.
The company also shipped DeepSeek Harness 0.1.1 the same day, with built-in support for the vision model. The harness is DeepSeek’s answer to Claude Code — an agentic coding tool that runs models in a development loop. Vision support means the harness can now read screenshots of error messages, UI layouts, and design files, not just parse text logs.
The broader picture
This release fits a pattern DeepSeek has followed all August: ship fast, price low, close capability gaps one at a time. V4 Pro went generally available on August 13. API prices increased 50–1,100 per cent but remained a fraction of Western competitors. Now vision lands as an experimental add-on to the cheapest model in the lineup.
The earlier V4-Flash public beta already undercut OpenAI’s discounted Luna by 30 per cent on input and 4x on output. Adding vision at no extra charge extends that moat. For New Zealand developers building AI agents — and there are more of them than the local market size would suggest — the practical question is whether an experimental model from a Chinese lab is reliable enough for production. The benchmarks say maybe. The price says it is worth finding out.
What stands out here is not that DeepSeek caught up to a frontier model. It did not, not quite. What stands out is how narrow the gap has become between a cheap experimental model and a $50-per-million-token flagship, on tasks that require both vision and agentic reasoning. A year ago, that gap was unbridgeable. Now it is twelve points on the hardest benchmark, and closing.
❓ FAQ
Is DeepSeek V4-Flash-Vision-Exp available now?
Yes. It is live on DeepSeek’s API platform as an experimental model. The model ID is deepseek-v4-flash-vision-exp. It supports JPEG, PNG, GIF, and WebP images.
How does it compare to Anthropic’s Opus 5? It does not, at least not in published benchmarks. DeepSeek compared against Opus-4.8, which is still supported but not Anthropic’s newest model. Nobody has published a head-to-head with Opus 5.
How much cheaper is it? V4-Flash costs roughly $0.22 per million input tokens at peak. Opus models cost around $50 per million. Images are billed at the same text rate, capped at 384 tokens per image.
Is it open-source? The V4-Flash text model weights are available. DeepSeek has not indicated whether the vision component will follow the same open-weight pattern.
Can it replace Claude for agent workflows? On most benchmarks it is within a few points. On repository-scale coding (NL2Repo) it trails by 12 points. For cost-sensitive workflows where near-frontier is good enough, it is a serious option. For the hardest coding tasks, the gap is still real.
📰 Sources
- The Next Web — DeepSeek launches experimental multimodal model to rival Anthropic
- DeepSeek on X — V4-Flash-Vision-Exp announcement
- explainx.ai — DeepSeek V4-Flash-Vision-Exp benchmarks and API details
- DeepSeek API Docs — Vision guide
- Bloomberg — DeepSeek unveils test model to rival Anthropic’s Opus 4.8
