DeepSeek V4 Flash Vision (Exp) API Pricing
Withdrawn on September 10, 2026, nineteen days after it surfaced. DeepSeek's change log says this model "has been retired", and the vision guide adds that the id is still accepted but answered by V4.1 Flash — which folded vision into the main Flash line and made a separate experimental endpoint unnecessary. It never reached GA, so there is no deprecation window to quote and nothing here was ever a commitment. While it ran it cost exactly what the text model cost: at peak $0.44/M input, $1.32/M output and $0.014/M cached input, half of each off-peak, on the same time-of-day card as V4 Flash. What made images cheap was not the rate but the ceiling: every image was resized to roughly 800x800 before inference, capping it at 384 tokens, so a 5000x5000 photo and a 2000x2000 one billed the same — about $0.00017 at peak. Its replacement resizes to roughly 1300x1300 instead, so the ceiling rose to 1024 tokens and a single image now costs up to $0.000307 at peak: more per image, on a card that is cheaper per token.
Run the numbers.
Calculator kept as a record, pre-loaded with the PEAK rates this endpoint charged until it was withdrawn — halve every result for an off-peak run. Images entered the same input bucket at up to 384 tokens each, so an image-heavy workload was priced by counting images rather than megabytes.
Real-world presets.
Repo-wide bug fix
Reading 100-page contracts
Reading a screenshot
Chart analysis turn
Counts use a chars-per-token calibration measured on the vendor's own published tokenizer (deepseek-ai/DeepSeek-V3.2-Exp, 2026-06-10). English prose is typically within a few percent; code and non-Latin scripts tokenize heavier. For billing-exact counts use the vendor's count-tokens API.
| Model | Input /M | Output /M | Effective blended | Context | Best for |
|---|---|---|---|---|---|
| DeepSeek V4 Flash Vision (Exp) Current | $0.44 cache $0.014 | $1.32 | $0.189 agentic 92/8 | 1M | What it charged while it ran |
| DeepSeek V4.1 Flash | $0.30 cache $0.006 | $1.20 | $0.15 −21% | 1M | GA vision on the main Flash line |
| DeepSeek V4 Flash | $0.44 cache $0.014 | $1.32 | $0.189 deprecated · same | 1M | Identical card, text only - also retired |
| DeepSeek V4 Pro | $1.32 cache $0.044 | $3.96 | $0.569 +3.0× | 1M | Harder reasoning, no image input |
| Qwen3 VL Flash | $0.05 | $0.40 | $0.078 deprecated · −2.4× | 256K | Cheapest vision tier, shorter window |
| GPT-5.6 Luna | $0.20 cache $0.02 | $1.20 | $0.152 −19% | 1.05M | Cheaper vision on a GA endpoint |
| Gemini 3.7 Flash | $0.75 cache $0.075 | $3.75 | $0.481 deprecated · +2.5× · promo | 1M | Vision plus audio and video input |
By comparison
Frequently asked.
What images actually cost on DeepSeek's experimental vision endpoint, and where the limits bite.
Q · 01 How much does one image cost? +
Q · 02 Does it cost more than the text model? +
$0.44 / $0.014 / $1.32 at peak and half that off-peak. There was no vision surcharge, no separate image unit and no minimum, so the only reason to pick V4 Flash instead was that this endpoint was experimental. Both were retired on the same day, and the model that replaced them keeps the arrangement: V4.1 Flash reads images on the same card as its text traffic.Q · 03 What does the -exp in the model id mean? +
DeepSeek-V4-Flash-Vision-Exp, so the experiment was baked into the id you called. We recorded it with preview status for that reason, and the reason held — DeepSeek withdrew it on September 10, 2026 with no deprecation window, nineteen days after our radar first saw it. That is what an -exp id is worth: treat one as a thing that can vanish between releases, and keep a GA endpoint in reach.Q · 04 How many images can one request carry? +
Q · 05 What does it give up against V4 Flash? +
Q · 06 Are images allowed anywhere in the conversation? +
400, and so does sending an image to any other DeepSeek model, which replies "This model does not support image". Supported formats are JPEG, PNG, GIF and WebP, detected from file content rather than from the file name.Q · 07 Which hours are peak? +
01:00-04:00 and 06:00-10:00 UTC, Monday through Friday. Every other hour — 17 of 24 on weekdays, and all weekend — is off-peak at exactly half the rate. Batch-style image work that can wait is the clearest case for scheduling: the same 600-image request costs about $0.05 off-peak instead of $0.10.