Previous state and the gap V4-Flash left

DeepSeek V4-Flash already offered text-based agent capabilities, a 1M-token context window, low API pricing, and OpenAI- and Anthropic-compatible API surfaces, but it did not accept image input. Teams needing vision had to route image workloads through another model or provider rather than the V4-Flash endpoint itself. That integration overhead is the slot V4-Flash-Vision-Exp is built to fill.

What V4-Flash-Vision-Exp changes

On 2026-08-21, DeepSeek released the experimental model deepseek-v4-flash-vision-exp on its API platform. It accepts image plus text input — JPEG, PNG, GIF, and WebP — through three paths: base64, an external URL, or the new Files API (launched the same day, free, with file_id reuse). Images are converted into tokens at a maximum of 384 tokens per image and billed as input tokens; inline and URL images are capped at 32 MiB, Files API uploads at 64 MiB, with a maximum of 8192 pixels per side and a detail=low mode that downscales to 512×512.

The model sits alongside the production deepseek-v4-flash and deepseek-v4-pro rather than replacing them. The API surface carries over: Chat Completions, the Anthropic-compatible Messages endpoint, and the Responses API all work; FIM Completion, supported on Flash and Pro, is not available here. Context length remains 1M tokens with a 384K max output, and the concurrency ceiling of 2500 matches V4-Flash.

Pricing and token economics

Pricing is identical to V4-Flash, re-verified against the official pricing page at 2026-08-28T09:49Z. Input tokens run $0.22 per million cache-miss off-peak and $0.44 peak; cache hits drop to $0.007 off-peak and $0.014 peak. Output is $0.66 per million off-peak and $1.32 peak. Peak hours are 01:00–04:00 and 06:00–10:00 UTC, Monday through Friday; all other hours are off-peak. Images billed under the 384-token-per-image rule fall into the same input-token structure, so a single image costs a small fraction of a cent at off-peak rates — but a request with many images still adds up linearly under that cap.

Vendor-reported capability claims

DeepSeek's release note makes three capability claims that should be read as vendor-reported, not independently verified. First, text capabilities — including agent behavior, reasoning, and world knowledge — are reported to match V4-Flash. Second, DeepSeek reports substantial multimodal-agent improvements over V4-Flash, but no independent validation was available at publication time. No benchmark table, methodology, or independent replication accompanied these claims in any artifact read for this analysis.

Limitations and unknowns

Several limits are load-bearing for any deployment decision. The model is officially labeled experimental; behavior, pricing surface, or availability may change, and the model could be withdrawn. FIM Completion is unsupported. DeepSeek's official release and documentation announce API access but do not document an open-weight release for this experimental checkpoint. DeepSeek released the model on 2026-08-21.

What to test before any migration

Can a team already using cheap text agents add vision at the same Flash token price without losing the accuracy or reliability its workflow requires? Vendor-reported claims do not settle that — especially for an experimental preview with no independent benchmark and no documented open-weight release. The defensible move is a structured test: take the screenshots, charts, document images, and multimodal tool-use workflows you actually run, exercise them against V4-Flash-Vision-Exp at the documented price points, and compare accuracy, latency, retries, task completion, and total workflow cost against the incumbent. Price alone is a reason to look; it is not, on its own, a reason to switch.

Frequently Asked Questions

What is deepseek-v4-flash-vision-exp? An experimental model on the DeepSeek API that adds native image input to the V4-Flash family. It was released 2026-08-21 and keeps the family's 1M-token context and pricing structure.

How much does it cost? Identical to V4-Flash per the official pricing page (re-verified 2026-08-28): $0.22 per 1M input tokens off-peak ($0.44 peak) and $0.66 per 1M output tokens off-peak ($1.32 peak). Images are converted into tokens and billed as input, at up to 384 tokens per image.

Can it replace your current vision model today? DeepSeek reports that multimodal-agent performance improves substantially over V4-Flash, but no independent validation was available at publication time. The defensible move is to benchmark it on your own workflow before replacing anything.

Are the model weights available? DeepSeek's official release and documentation announce API access but do not document an open-weight release for this experimental checkpoint.