DeepSeek V4 Flash Vision Adds Weekend Off-Peak Pricing: The Coding Agent Cost Math

DeepSeek has added deepseek-v4-flash-vision-exp to its official API rate card at the same token rates as V4 Flash, while a billing rule effective August 23 makes every Saturday and Sunday in Beijing time off-peak. For coding agents that inspect screenshots or can defer bulk work, both changes matter.

The official DeepSeek pricing page lists the experimental vision model with a 1M-token context window, image and text input, tool calls, and the Responses and Anthropic-compatible APIs. Images are converted to tokens according to their dimensions and billed with text input. That means there is no separate per-image fee to hide in a calculator, but image resolution directly affects input-token volume.

The Rates That Apply

Off-peak V4 Flash Vision costs $0.22 per million cache-miss input tokens, $0.007 per million cached input tokens, and $0.66 per million output tokens. Peak rates double to $0.44, $0.014, and $1.32. The ordinary V4 Flash endpoint uses the same rate card. V4 Pro remains three times the Flash rates, so a visual executor can sit beneath a Pro planner without forcing every screenshot through the premium model.

DeepSeek still defines weekday peak blocks as 01:00–04:00 and 06:00–10:00 UTC. The new exception is simpler: Saturdays and Sundays in Beijing time receive off-peak rates all day. Because Beijing time is UTC+8, teams should convert the weekend boundary before scheduling a Friday-night job from Europe or North America.

A Browser-Test Example

Assume a visual regression agent handles 100 test cases. Each case sends 40,000 tokens of text and image input and produces 8,000 output tokens across inspection, diagnosis, and a suggested patch. The batch totals 4M input and 0.8M output tokens. At weekend off-peak rates, uncached spend is $0.88 plus $0.528, or $1.408. At peak, the same work costs $2.816.

Caching changes the result further. If 70% of the input is a stable design system, test instructions, and repeated page context, only 1.2M tokens pay the $0.22 cache-miss rate while 2.8M pay $0.007. Input cost falls from $0.88 to about $0.284. Add output and the weekend batch is about $0.812. The savings come from preserving a byte-identical prefix, not merely sending similar instructions.

Where Vision Saves Money

• Delegate visual verification. Keep architectural reasoning on a stronger model, then send screenshots and bounded acceptance criteria to Flash Vision.
• Resize deliberately. A full-resolution desktop capture may contain far more billable visual tokens than a cropped component image.
• Schedule deterministic suites. Nightly and weekend visual checks rarely justify paying weekday peak rates.
• Cache the stable frame. Put policies, component catalogs, and test conventions before dynamic screenshots and errors.

Do Not Budget an Experimental Endpoint Like GA

The endpoint carries an experimental suffix. Budget for alias changes, capability changes, and a fallback path. Store the model ID and rate schedule in configuration, record image dimensions with usage, and test the same suite against a second provider. A cheap run that blocks release because the endpoint changed is not cheap in outcome terms.

Also separate model cost from browser infrastructure. Screenshots require a browser or device session, storage, network traffic, and sometimes paid testing infrastructure. The token invoice above is only the inference layer. The useful KPI is total cost per accepted visual fix: model tokens plus runtime plus retries plus review, divided by fixes that pass the full suite.

Build a Weekend Queue, Not a Weekend Assumption

Tag jobs as interactive, deferrable, or deadline-bound. A scheduler can hold only deferrable visual sweeps for the Beijing-time weekend and release them with a maximum concurrency limit. If the queue is too large, waiting can delay feedback until after developers have changed the same interface again. Put an expiry on queued work so an obsolete screenshot suite does not run merely because cheap hours arrived.

Export billing-time timestamps rather than relying on submission time. A job launched off-peak can continue into a peak block, and a provider may assign usage by processing time. Compare the invoice with your request logs for one week before forecasting a perfect 50% schedule discount. Record timezone conversion in code and tests; daylight-saving changes elsewhere must not move the Beijing-time rule.

Evaluate Vision Quality by Failure Type

Create a small evaluation set for clipped text, low contrast, responsive breakpoints, modal focus, chart labels, and dense tables. Count false alarms separately from missed regressions. A model that catches more cosmetic differences may consume reviewer time without protecting releases. Route ambiguous or high-impact cases to a stronger verifier instead of increasing reasoning on every screenshot.

Keep the original screenshot, crop coordinates, model response, proposed patch, and final test result under one run ID. That evidence shows whether cost increases came from larger images, more agent turns, or lower first-pass accuracy. It also makes migration away from the experimental endpoint measurable rather than anecdotal.

Bottom Line

DeepSeek has made a practical low-cost visual executor available and given schedulers a predictable weekend discount. Use the model for bounded screenshot analysis, protect cacheable prefixes, crop inputs, and move noninteractive batches into the Beijing-time weekend. Then compare accepted fixes rather than headline token rates in the AI Cost Calculator.