Eyes for the cheap Flash
On August 21, DeepSeek quietly added a new model to its API: V4-Flash-Vision-Exp.
The name is long and tagged experimental. But what it does is simple: it gives the July 31 V4-Flash, the cheap edition, the ability to finally see.
Not a new model, a vision patch
Let us be clear: this is not another big model trained from scratch.
It is the original V4-Flash with vision bolted on. The pure-text half, agent work, reasoning, world knowledge, is identical to the base, not a step down. What is new is reading images, screenshots, and charts.
For agent builders, that means your bot can finally look at the screen. Until now you needed to step up to a premium model like Opus 4.8 for that. The cheap Flash can now vaguely manage it.
Zoom out and the move is bigger than it looks. Through August 2026, labs have quietly started adding vision to their cheap tiers, not to show off pretty generated pictures, but so agents can read screens, charts, and receipts. An agent that can see is the one that can finally take over the repetitive back-office work nobody wants. DeepSeek is riding that exact wave.
Why vision matters for agents
The point is not pretty pictures. It is screen reading.
An agent that can read a screenshot, a chart, or a dashboard can do the boring back-office jobs people hate: check a status page, read a receipt, parse a form, navigate a GUI it has never seen. Vision turns a text-only bot into something that operates in the messy visual world most enterprise software still lives in. DeepSeek says the text-agent benchmarks sit right next to Flash, Terminal Bench 2.1 at 83.9 versus 82.7 and DeepSWE 59.3 versus 54.4, small deltas. The vision addition is the story, not a smarter text model.
The numbers, translated
The context window is 1M tokens with 384K max output, matching Flash and Pro, no shrinkage.
Images are billed as input tokens, capped at 384 each. A single product screenshot costs almost nothing. But push 600 images and hit the 64MB image ceiling and you reach the boundary. Do not assume seeing is free and flood it.
Pricing: same as Flash, but the peak bites
Pricing is identical to the original V4-Flash.
Peak hours run $0.44 / $1.32 per million input / output tokens, and off-peak halves that. Looks generous, right? The catch is when the peak sits: 01:00 to 04:00 and 06:00 to 10:00 UTC. In Beijing time that is 09:00 to 18:00.
The eight hours you work hardest are exactly the eight hours it costs most.
Remember this, or your month-end bill will confuse you.
The caveat: approaching, not beating
DeepSeek published a string of numbers: across 11 disclosed benchmarks, V4-Flash-Vision-Exp beats Opus 4.8 on 3, including Agents Last Exam at 27.3 versus 25.7.
But approaching Opus 4.8 is far more honest than beating it.
On the hardest repository-level coding task, NL2Repo, it scores 57.7 against Opus 4.8's 69.7, a full 12-point gap that DeepSeek does not try to disguise. More importantly, every figure is vendor-reported with no independent audit; trust it at your own discount. And it is API-only, no downloadable weights, so teams wanting local deployment or their own fine-tune should wait.
Being able to see is not the same as seeing well. An agent needs to act correctly after reading the screen, and that is the real divide.
A free Files API ships alongside: upload an image once, reference it by file_id, reuse it across requests without re-uploading. DeepSeek Harness 0.1.1 supports the new model out of the box, so wiring it into an agent framework takes minutes, not days. Concrete case: a seller wants a bot that opens a product page, reads the screenshot, and drafts the listing copy. Before this model that meant paying Opus 4.8 rates. Now the cheap Flash can read the screenshot and draft. The copy still needs a human eye, but the first pass is nearly free.
Why this matters to you
If you already run on DeepSeek's API and want an agent that reads a screenshot or a chart, this is testable today, no provider switch, no price premium. Why not.
But for serious visual coding that reasons across a full repository, Opus 4.8 and GPT-5.6 Sol still lead. Lightweight screen-reading agents: go now. Heavy visual coding: wait.
My read: it is a competent eyes add-on, not an end-to-end visual coding engine. Set the expectation right and you will not be disappointed.