DeepSeek quietly shipped deepseek-v4-flash-vision-exp today — a vision-capable build of V4-Flash priced identically to the text-only version. 1M context, 384K output, 384 tokens per image, 600 images per request, OpenAI and Anthropic API compatibility. It is the first V4 vision model. It sits at #1 on Hacker News with 260 points. The cheap-tier vision story is over.

DeepSeek Just Gave V4-Flash Eyes. The $0.66/M-Output Vision Model Is The Most Consequential API Drop Of August.

Hey guys, Mr. Technology here.

It is Thursday, August 21, 2026, and DeepSeek shipped deepseek-v4-flash-vision-exp this morning without a blog post, without a press release, without a model card on Hugging Face — by quietly appending a third model name to the public DeepSeek API documentation. The Hacker News submission hit 260 points and 66 comments within three hours. The reason it matters: it is the first vision-capable model in the V4 family, priced identically to the text-only V4-Flash, with 1M context and OpenAI/Anthropic API compatibility from launch. Cheap-tier vision just became real. Read that as many times as you need to.

If you are routing agentic traffic through any vision-capable API right now, this is the week you start writing the migration plan. If you are not, this is the week you start thinking about what becomes possible when a vision-capable frontier-tier model costs less than your coffee.

What Shipped On August 21

deepseek-v4-flash-vision-exp — a vision-capable experimental build of V4-Flash. Same MoE architecture, same DSpark speculative decoder attached, same 304B total parameters, same post-training lineage. The change is a new visual encoder bolted onto the same Flash checkpoint and a new vision capability flag on the API.
API name to call it: deepseek-v4-flash-vision-exp. Replaces nothing — it is additive to the V4 line.
Pricing: identical to V4-Flash-0731. Peak $1.32/M input (cache miss) and $1.32/M output. Off-peak $0.22/M input (cache miss) and $0.66/M output. Cache hit: $0.014/M peak, $0.007/M off-peak.
1M token context, 384K max output, same as the rest of V4.
Both OpenAI-compatible and Anthropic-compatible endpoints from launch. Base URL https://api.deepseek.com for OpenAI, https://api.deepseek.com/anthropic for Anthropic SDK users.
Concurrency limit 2,500 per account, same as V4-Flash.

What The Vision Part Actually Does

From the DeepSeek Vision guide, the model accepts image inputs alongside text via three methods:

1. Base64 inline — encoded JPEG/PNG/GIF/WebP up to 32 MiB, body size limit 48 MiB. 2. External HTTP(S) URL — model downloads, max 8,192-character URL, max 32 MiB image, 60-second download timeout. 3. Files API reference — upload once via file-api-* ID, reuse across requests, up to 64 MiB per image.

Per-request limits: 600 images per request, max 64 MiB total image payload (or 200 MiB if you use file_id), max 8,192 px per side (drops to 4,096 px when you hit 15+ images).

The interesting constraint is the image tokenizer. Images are resized so the total pixel count after resize is roughly that of an 800×800 image. Upper bound: 384 tokens per image, regardless of input resolution. A 2,000×2,000 image and a 5,000×5,000 image consume the same number of tokens after resize. Multi-image requests are billed per-image independently, no bulk discount.

That detail matters. The resolution cap is the constraint every agent harness is going to bump into this week. For OCR of an A4 page at 300 DPI, 800×800 is below the threshold to read small text reliably. For screenshots of a webpage, it is fine. For plots and matplotlib figures, it is fine. For receipt OCR at a checkout counter, you will need a crop-and-zoom harness on top.

The detail level flag helps: detail: "low" downscales to 512×512, detail: "high" keeps the original, detail: "auto" currently equals original. The Flash-Vision-Exp docs hint that the eventual Pro vision build will raise the resolution cap.

Why This Is Bigger Than It Looks

Three reasons this is the most consequential API drop of August 2026.

Reason one: Liang Wenfeng explicitly committed to multimodal in July. The leaked investor-meeting transcript had DeepSeek's founder explicitly walking back his earlier text-only-is-sufficient-for-AGI stance: "Multimodal approaches ultimately need to be implemented. As a component, we will undoubtedly implement multimodal support — and we are already doing so." The V4-Flash-Vision-Exp is the first public deliverable against that commitment, three weeks after the meeting, on the cheap tier. The Pro vision build is implied to be next, probably within a month.

Reason two: it removes the open-weights vision gap at the cheap tier. Until today, the only vision-capable models you could route agentic traffic through at frontier quality were GPT-5.6 Vision, Sonnet 5 with vision, Gemini 3.6 Flash, Claude Opus 4.8 with vision, and a handful of open-weights specialist VLMs (Qwen3-VL, InternVL 3.5, Llama 3.3-Vision, MiniMax-H3) that each had meaningful capability gaps. The DeepSeek V4-Flash-Vision-Exp inherits V4-Flash-0731's agentic coding and tool-use capability — the same model that ranks ahead of a 428B proprietary model on Artificial Analysis — with vision grafted on. The capability floor for cheap-tier vision just lifted.

Reason three: the pricing curve inverts. V4-Flash-Vision-Exp at off-peak $0.66/M output is roughly 1/23rd the price of Opus 4.8 with vision ($15/M output). Against Sonnet 5 with vision ($15/M output) it is 1/23rd. Against GPT-5.6 Vision ($10/M output) it is 1/15th. For an agent harness that processes screenshots on every loop iteration — a Playwright/Chrome DevTools Protocol loop, a desktop control loop, a 3D-printer watch loop — that is not a marginal cost reduction. That is the difference between "we can ship it" and "we cannot afford it."

What The HN Thread Got Right

The top-of-thread reactions hit three notes that match my own read:

The 800×800 cap will drive a crop-and-zoom harness pattern. One commenter captured it well: "For OCR of small text and the like a zoom ability should be sufficient." Expect vLLM-side tooling, LangChain vision loaders, and DeepSeek Harness itself to ship a crop tool next to the Files API by end of September.
V4-Flash-0731 has been hallucinating vision capabilities for weeks. A long thread on the previous build noted that 0731 frequently assumed it could see and would invent text-based image-analysis tools when it could not. "An upgrade for the model," one commenter called it. The hallucination surface is now closed.
The use cases people actually care about are screenshots. Frontend dev with headless browser screenshots, OCR of paper documents, processing graphs and charts, matplotlib figure self-correction, design sketches and wireframes, handwriting recognition, Figma-via-CDP. "Surveillance is the obvious one," said one HN comment. The first reply: "Frontend Dev? I do not really understand" — followed by half the thread explaining that, yes, you point your agent at a browser, take a screenshot, and the model fixes its own UI mistakes.

The model fails the simple clock test in informal HN runs — one user reported it answered 5:10 with the second hand on the 9 (45 seconds) when the answer was 8:09:25. Compare that to Qwen3.8 27B which got 8:10 with the second hand on the 5 (25 seconds). For a cheap-tier vision model, close-but-wrong is the calibration zone. Worth the price.

What To Ship This Week

1. Wire the Files API path first. Inline base64 will burn your request body size limit on the first document-processing workload. Uploading via file-api-* and referencing the ID is the production path. 2. Build a screenshot-driven agent loop on a non-vision harness first. Same image, same tool calls, same harness. Switch the model name from deepseek-v4-flash to deepseek-v4-flash-vision-exp and verify each step still passes. The OpenAI and Anthropic SDKs both accept the model name with no code change. 3. Set detail: "low" for screenshots at scale. 512×512 is enough for layout detection, button presence, text legibility at body-text sizes. For OCR of dense forms, default to detail: "auto". 4. Budget for the Pro vision build next. The 384-tokens-per-image cap is the headline constraint. DeepSeek's own docs hint at a higher resolution path in the Pro vision tier. If your workload genuinely needs higher resolution, do not invest too much in Flash-Vision-Exp's resolution ceiling — wait four weeks. 5. Re-tier vision in your routing layer. Sonnet 5 with vision moves from "default cheap vision" to "fallback for cases where the Flash-Vision-Exp resolution cap is binding." GPT-5.6 Vision moves to "default vision for cases where OpenAI-only tool integration matters." Opus 4.8 with vision stays as "the long-tail vision tier where the few points of capability gap are worth the 23x price."

Mr. Tech's Take

DeepSeek's V4 line is now in its third weekly update in three weeks: V4-Flash-0731 on July 31, V4-Pro-0813 on August 13, V4-Flash-Vision-Exp today. Each release has shipped a feature that would have been the headline release of any given month in 2024 or 2025: cheap-tier agent king, MIT-licensed 1.7T frontier, vision at $0.66/M output. The cadence is now weekly at DeepSeek. The frontier is moving at a cadence the rest of the field has not matched.

The 800×800 vision cap is the constraint. The 1M context is the enabler. The Anthropic-API compatibility is the migration path. The pricing is the procurement event.

If you are building an agentic product on top of a vision API right now, you have until the end of next week to either (1) build the crop-and-zoom harness on top of Flash-Vision-Exp, or (2) budget to wait for the Pro vision build that raises the resolution cap. There is no version of this story where you stay on Sonnet-5-with-vision as your default for high-volume vision traffic and stay competitive on cost. The 23x multiplier compounds every month.

The cheap-tier vision model is now the cheap-tier vision model that thinks.

— Mr. Technology

Sources: