DeepSeek V4 API adds experimental Flash Vision model

What changed with DeepSeek V4 Flash Vision, V4 Pro pricing and Harness?

DeepSeek has added the experimental V4 Flash Vision model to its API and current pricing table. It accepts text plus images, uses the model identifier deepseek-v4-flash-vision-exp, and is supported by the Harness 0.1.1 release-candidate line. The same DeepSeek V4 page continues to track V4 Pro 0813 and current peak/off-peak pricing. Developers should treat Vision, Pro and the prerelease Harness as separate workflow choices and verify current prices before use. Leon for The AI Guy owns this researched update and its next source refresh.

DeepSeek V4 Flash Vision is live as an experimental API model

DeepSeek's current documentation lists DeepSeek-V4-Flash-Vision-Exp under the API identifier deepseek-v4-flash-vision-exp. The experimental model accepts images with text and supports JPEG, PNG, GIF and WebP through inline base64, public image URLs or the Files API. DeepSeek lists a 1 million-token context window, up to 384,000 output tokens, JSON output, tool calls, Responses API and Anthropic API support; FIM completion is not supported. These are vendor specifications. The AI Guy has not tested the model, its image understanding, endpoints or limits.

How Vision, V4 Pro and Harness differ

V4 Flash Vision is the experimental image-capable API model. V4 Pro remains a separate text model with its own deepseek-v4-pro identifier and higher posted rates. DeepSeek Harness 0.1.1 rc.2 is a release-candidate integration path that adds the Vision adapter, reusable Files API uploads and automatic image preprocessing. This separation matters: a model capability, a token price and a prerelease agent interface are three different workflow decisions, and none establishes production readiness.

What changed in the model listing

DeepSeek's official API documentation now identifies the Pro model version as DeepSeek-V4-Pro-0813. The documented API identifier remains deepseek-v4-pro. This does not establish how third-party providers, pinned deployments or cached configurations will handle the version change. South China Morning Post independently describes 0813 as an update to April's V4 Pro preview, not a new model family.

DeepSeek Harness adds Vision support in a prerelease

DeepSeek Harness remains an MIT-licensed developer preview that warns of compatibility-breaking changes. Its 0.1.1 release-candidate line adds the DeepSeek-V4-Flash-Vision-Exp adapter; current rc.2 prioritizes Files API uploads for reusable images and adds automatic resizing and format conversion. DeepSeek still documents npx @deepseek-ai/dsh web as the local start path. This is an integration route, not evidence of stability, security, provider compatibility or production readiness. The AI Guy has not installed or tested the Harness release.

The documented context limits

The official page lists a 1 million token context window and a maximum output of 384,000 tokens for V4 Pro. These are vendor specifications; The AI Guy has not tested long-context quality or compatibility across third-party tools.

Vision input is billed as tokens

DeepSeek says image inputs are converted into tokens based on their dimensions and billed together with text input. Its current table lists the Vision model at $0.007 off-peak or $0.014 peak per million cached input tokens, $0.22 or $0.44 per million uncached input tokens, and $0.66 or $1.32 per million output tokens. On weekdays, peak hours remain 01:00–04:00 and 06:00–10:00 UTC. From 00:00 Beijing time on August 23, DeepSeek says off-peak rates apply throughout Saturdays and Sundays measured in Beijing time. Prices are time-sensitive and should be rechecked before budgeting.

Current V4 Pro off-peak prices

DeepSeek's current table lists V4 Pro off-peak rates of $0.022 per million cached input tokens, $0.66 per million uncached input tokens and $1.98 per million output tokens. V4 Pro is a separate model from V4 Flash Vision, which has lower posted rates.

Current V4 Pro peak prices and time windows

DeepSeek's current V4 Pro peak rates are $0.044 per million cached input tokens, $1.32 per million uncached input tokens and $3.96 per million output tokens. On weekdays, peak windows are 01:00–04:00 UTC and 06:00–10:00 UTC; the other weekday hours are off-peak. Effective 00:00 Beijing time on Sunday, August 23, DeepSeek says off-peak rates apply throughout Saturdays and Sundays measured in Beijing time. Recheck the official pricing page before budgeting because DeepSeek says prices and billing rules may change.

What this means for agent workflows

A practical cost check is to separate cached input, uncached input, image-derived input and output, then compare the correct model and time window. For weekday planning, the posted peak windows still apply. From 00:00 Beijing time on August 23 — 16:00 UTC on August 22 by timezone conversion — DeepSeek says off-peak rates apply throughout Saturdays and Sundays measured in Beijing time. Builders should use the Beijing calendar rather than assume a UTC weekend boundary. Flexible jobs may be candidates for off-peak scheduling, but that is a planning framework rather than a savings, quality or reliability guarantee. The AI Guy has not tested DeepSeek's billing or cache behavior.

The bottom line

DeepSeek now offers an experimental image-capable V4 Flash API model alongside V4 Flash and V4 Pro. The Vision model has posted token prices and a documented Harness prerelease path, while V4 Pro retains its separate model identifier and higher pricing. Treat the models and Harness maturity separately, recheck the current table, and validate the workflow before production use.

Related tools and pages

DeepSeek researched profileOpen related page AI LLM and chatbot toolsOpen related page Grok 4.6 long-running agentsOpen related page

Official sources

DeepSeek official Models & Pricing documentation ↗Checked 2026-08-22
South China Morning Post independent report ↗Checked 2026-08-13
DeepSeek official Harness repository ↗Checked 2026-08-14
TechNode independent Harness report ↗Checked 2026-08-14
DeepSeek official Vision guide ↗Checked 2026-08-21
DeepSeek official Flash Vision announcement ↗Checked 2026-08-21
DeepSeek Harness 0.1.1 rc.2 release ↗Checked 2026-08-21
The Next Web independent Flash Vision report ↗Checked 2026-08-21

This is researched analysis based on public product information. The AI Guy has not independently benchmarked the feature described here.

FAQ

What is DeepSeek-V4-Pro-0813?

It is the model version currently listed for the deepseek-v4-pro API identifier and an update to the V4 Pro preview released in April.

When did DeepSeek's peak and off-peak API prices take effect?

DeepSeek's pricing transition took effect at 16:00 UTC on August 16, 2026. The current official table retains the peak and off-peak rates.

What are the DeepSeek V4 Pro peak windows?

On weekdays, the official page lists 01:00–04:00 UTC and 06:00–10:00 UTC as peak windows. Effective 00:00 Beijing time on August 23, 2026, off-peak rates apply throughout Saturdays and Sundays measured in Beijing time.

What is DeepSeek Harness?

DeepSeek describes it as an open-source agent harness with a plugin-based architecture. It remains a developer preview, and its current 0.1.1 line is a release candidate that warns of compatibility-breaking changes.

Is DeepSeek V4 Flash Vision available through the API?

Yes. DeepSeek lists the experimental model under deepseek-v4-flash-vision-exp. The AI Guy has not tested access, limits or output quality.

How are images priced in DeepSeek V4 Flash Vision?

DeepSeek says images are converted into tokens based on their dimensions and billed together with text input. Its published token rates are time-sensitive and should be rechecked before use.