DeepSeek V4 Flash Vision (Exp) vs DeepSeek V4 Flash: Same Price, One Sees

DeepSeek V4 Flash Vision (Exp) and DeepSeek V4 Flash are the same model with exactly one difference, and that difference is the entire story: the vision model sees images and the text model does not. Launched today, August 21, 2026, as an experimental release, DeepSeek V4 Flash Vision (Exp) carries DeepSeek V4 Flash's text brain unchanged — the same 1M-token context, the same 384K max output, the same token prices — and adds image input that rearranges its scores on the agent benchmarks where the text model quietly fails. The only real question is whether "experimental" costs you more than it saves.

The two models

DeepSeek V4 Flash is the company's official budget workhorse: a Mixture-of-Experts model with roughly 284B total parameters and 13B active, text-only, optimized for high-concurrency everyday workloads, with a 2,500-token-per-second concurrency budget on the vendor's API. DeepSeek V4 Flash Vision (Exp) is the same weights family with a vision encoder in front of it, billed identically, and marked experimental. Everything below the eyes is shared.

• Parameters — Vision-Exp: 284B total / 13B active MoE vs V4-Flash: 284B total / 13B active MoE (same family)

• Input — Vision-Exp: text + images (JPEG, PNG, GIF, WebP) vs V4-Flash: text only

• Context — Vision-Exp: 1M tokens vs V4-Flash: 1M tokens

• Max output — Vision-Exp: 384K tokens vs V4-Flash: 384K tokens

• Thinking modes — both support thinking and non-thinking

• API formats — both: Chat Completions, Messages, Responses

• Price — identical (both bill at V4-Flash's peak/off-peak token rates)

• Status — Vision-Exp: experimental, no GA date vs V4-Flash: official release (V4-Flash-0731)

Where vision changes the scoreboard

On pure text, Deep​Seek says the two are on par, and there is no reason to doubt it — the vision model is built from the same text capability. The divergence shows up on agent benchmarks that contain screenshots, charts, and UI images, where the text-only model literally ignores the multimodal content. All figures below are vendor-reported from today's launch note, not independently reproduced:

• ApexBench Pass@1 — Vision-Exp 36.5 vs V4-Flash 26.2

• Agents' Last Exam — Vision-Exp 27.3 vs V4-Flash 25.2

• Terminal-Bench 2.1 — Vision-Exp 83.9 vs V4-Flash 82.7

• NL2Repo — Vision-Exp 57.7 vs V4-Flash 54.2

• DeepSWE — Vision-Exp 59.3 vs V4-Flash 54.4

• Chartography — Vision-Exp 64.3 (V4-Flash cannot attempt)

• ZeroBench Pass@5 — Vision-Exp 35.0 (V4-Flash cannot attempt)

The single biggest jump is ApexBench, where adding vision lifts the score from 26.2 to 36.5 — a ten-point swing that is not the model thinking harder, but finally reading the task. For an agent that operates through a screen — a browser, a desktop, a terminal with rendered output — the text model was flying blind on exactly the parts of the task that carry the information.

The price question has one answer

There is no price difference, and this is where the comparison becomes a no-brainer in one direction. DeepSeek V4 Flash Vision (Exp) bills at exactly the same rates as DeepSeek V4 Flash, which since August 16, 2026, have been peak/off-peak:

• Input, cache miss — $0.22 per 1M tokens off-peak, $0.44 peak (both models)

• Input, cache hit — $0.007 per 1M tokens off-peak, $0.014 peak (both models)

• Output — $0.66 per 1M tokens off-peak, $1.32 peak (both models)

• Peak hours — 01:00–04:00 and 06:00–10:00 UTC, both models

The only added cost vision brings is the image itself: each image is tokenized at up to 384 tokens, so a screenshot costs roughly 0.0085 cents off-peak. Even a vision-heavy agent reading 50 images per run adds under a cent of input. Choosing the text model to save money is choosing pennies over eyesight — the saving is not real.

When to pick which

Choose DeepSeek V4 Flash Vision (Exp) if your pipeline can ever receive an image. UI-testing and screenshot-based QA agents, web-browsing agents that need to read the page they are on, chart and diagram extraction, document screenshots, error-message captures from a user's screen — in every one of those, the vision model is the strictly better model at the same price. There is no text-quality penalty, per the vendor, so the only reason not to use it is stability.

Choose DeepSeek V4 Flash if your traffic is pure text and nothing about it will change. For coding completion, retrieval, and text-only agent loops, the two models score the same on text and the official release is the safer bet by definition. If you are already on V4-Flash and never send an image, there is no reason to switch — and no reason not to have the option in your routing config either.

The one real difference: experimental status

Everything above the eyes is identical; the cost of vision is negligible; the benchmarks favor the vision model. The single factor that can legitimately keep you on DeepSeek V4 Flash in production is that the vision model is experimental. Deep​Seek labels it so, gives no GA date, and says it is not recommended for production use. The text brain behind it is proven, but the vision path is new, and an experimental API surface is exactly where you do not want a silent failure at 2 AM.

This is the one place the comparison is not settled by specs, and it is also the one place a router changes the answer. On OrcaRouter both models are live — deepseek/deepseek-v4-flash-vision-exp and deepseek/deepseek-v4-flash — behind a single API key at the provider's list price, passed through with zero markup. Because the same platform routes both, you can adopt the vision model where it earns its keep and pin the rest to the official release, with automatic failover so an experimental hiccup never takes the whole pipeline down. You get the ten-point ApexBench gain on the calls that need it and the official release's stability on the calls that do not, without a second contract or a second code path.

Bottom line

If your workload can receive images, DeepSeek V4 Flash Vision (Exp) beats DeepSeek V4 Flash on every axis that matters and loses only on the one axis you can engineer around: experimental status. If your workload is pure text, the models are equal on output and the official release wins on stability. The two are not really competitors — the vision model is the text model with a skill unlocked, at no extra charge. The only decision worth making is how much of your traffic you are willing to point at an experimental API, and that is a routing decision, not a model decision.