This guide uses source checks from Sep 2, 2026. Provider and gateway prices can change; preserve the checked date with every forecast.
Why This Topic Matters Now
DeepSeek V4 content has mostly focused on text routing, long context, and Pro versus Flash planning. The official DeepSeek pricing page checked on Sep 2, 2026 adds a more specific production question: V4 Flash Vision Exp appears beside V4 Flash and V4 Pro, with image inputs converted into tokens and billed as input tokens. A multimodal trial cannot borrow a text-only cost worksheet and stay accurate.
This guide is written for Tier 1 and Tier 2 teams testing screenshot review, document images, visual QA, multimodal support, or design-agent workflows. It uses official DeepSeek docs checked on Sep 2, 2026 and AIWave live route data checked the same day. The rule is practical: treat image input, text input, cache class, output length, peak window, and route availability as separate fields before using a multimodal route in production.
Source Facts Checked Today
DeepSeek pricing checked on Sep 2, 2026 lists deepseek-v4-flash, deepseek-v4-pro, and deepseek-v4-flash-vision-exp. All three show OpenAI-format and Anthropic-format base URLs, 1M context, 384K maximum output, JSON output, tool calls, Responses API support, Anthropic API support, and chat prefix completion. FIM completion is listed for non-thinking mode on text routes and not supported for the vision-exp row.
The official price rows checked on Sep 2, 2026 list V4 Flash Vision Exp with the same token prices as V4 Flash: cache-hit input $0.007 off-peak and $0.014 peak, cache-miss input $0.22 off-peak and $0.44 peak, and output $0.66 off-peak and $1.32 peak per 1M tokens. Peak windows are 01:00-04:00 and 06:00-10:00 UTC, Monday through Friday, and all other hours are off-peak. The page says images sent to V4 Flash Vision Exp are converted into tokens based on dimensions and billed with text input tokens.
AIWave /api/pricing checked on Sep 2, 2026 returned 63 route records and pricing_version a42d372ccf0b5dd13ecf71203521f9d2. Parsed text-route examples before account-group math included DeepSeek V4 Flash at $0.638 input, $1.914 output, and $0.020288 cache-hit input per 1M tokens, and DeepSeek V4 Pro at $1.914 input, $5.742 output, and $0.063736 cache-hit input. For a vision-exp trial through any gateway, confirm actual route availability and feature support before copying direct-provider assumptions.
Planning Matrix
A source-dated planning matrix keeps the page useful for engineers and procurement reviewers. It turns a search query into an auditable route decision instead of a loose model preference.
| Trial field | Text-only mistake | Multimodal control |
|---|---|---|
| image_tokens | Screenshots hidden inside prompt notes | Record converted image input tokens |
| image_dimensions | No way to reproduce token count | Store width, height, and count |
| cache_class | Assumes screenshot reuse behaves like text | Measure cache-hit input separately |
| peak_window | Ignores weekday UTC price window | Store request time and timezone |
| output_cap | Visual analysis writes long reports | Set max tokens per task type |
| route_available | Assumes gateway exposes the direct route | Check models endpoint before launch |
Implementation Pattern
The implementation pattern keeps credentials as placeholders, pins the AIWave base URL, records the model, and leaves room for route-specific controls. Production applications should move credentials into environment or secret storage.
from openai import OpenAI
client = OpenAI(api_key="YOUR_API_KEY_HERE", base_url="https://aiwave.live/v1")
def record_multimodal_probe(model: str, image_url: str, instruction: str):
response = client.chat.completions.create(
model=model,
messages=[{
"role": "user",
"content": [
{"type": "text", "text": instruction},
{"type": "image_url", "image_url": {"url": image_url}},
],
}],
max_tokens=600,
temperature=0.1,
)
return {
"model": response.model,
"usage": response.usage.model_dump() if response.usage else None,
"source_checked_at": "2026-09-02",
}
Do Not Reuse a Text-Only Worksheet
A text worksheet usually tracks input tokens, output tokens, model, and maybe cache. A multimodal worksheet needs more fields: image count, dimensions, conversion rule source, text input, cache class, output cap, request timestamp, peak-window status, and route availability. If the team cannot reproduce how image input became billed input tokens, the forecast is not ready for procurement review.
Confirm Route Availability First
DeepSeek's official page can list a direct route before a gateway exposes the same feature. An AIWave evaluator should call the models endpoint, check the docs, or ask support before building a trial around V4 Flash Vision Exp. If only text DeepSeek routes are visible, run the multimodal experiment directly or postpone the gateway test. Do not infer feature support from a similar text route name.
Track Image Dimensions
DeepSeek says image inputs are converted into tokens based on dimensions. Store width, height, image count, and source document type beside the request. A screenshot-heavy workflow can look small in request count while carrying large input. If the application downscales or crops images, record that transformation too. Later cost reviews should be able to tell whether spend moved because volume grew or image sizes changed.
Handle Peak Windows
The official DeepSeek page uses weekday UTC peak windows. US, UK, Germany, Japan, and Singapore teams should store request time in UTC and local time so they can understand why identical prompts cost different amounts on direct DeepSeek. AIWave's public DeepSeek text rows are all-day gateway examples, but direct-provider comparisons should still preserve peak status. Keep those two concepts separate in the ledger.
Cap Visual Output
Visual analysis can produce long explanations, especially for UI QA, document review, and design critique. Set task-specific max tokens before the first run. A quick screenshot label may need a short cap; a compliance review may need a longer cap and a human reviewer. Without output budgets, a multimodal trial can look unpredictable even when image input is well measured.
Use Internal Links for DeepSeek Searchers
Readers evaluating DeepSeek should move into Models docs, Chat Completions, Pricing, Predictable Pricing, the DeepSeek peak-window guide, and the DeepSeek Responses acceptance-test guide. That cluster keeps vision budgeting connected to existing text-route evidence.
Procurement Review
Procurement should ask for official DeepSeek source URL, checked date, route name, image-token conversion source, peak-window status, output cap, AIWave pricing_version when a gateway is used, and one redacted usage object. Engineering should show at least one small image, one large image, and one text-only control. That comparison makes it possible to explain the marginal cost of vision.
Final Checklist
A DeepSeek V4 Flash Vision Exp trial is ready when the team has official source dates, image dimensions, image-token fields, text-token fields, peak-window status, output caps, route availability evidence, and a redacted usage ledger. Recheck DeepSeek and AIWave before launch; a direct-provider vision row should never be treated as automatic gateway support.