DeepSeek shipped V4-Flash-Vision-Exp, a multimodal model that reads screenshots, charts, and dashboards — at the same price as the text-only flash. It's a striking step: the cheap, fast tier now does what premium models did a year ago.
What it is
deepseek-v4-flash-vision-exp is DeepSeek's first multimodal model in the V4 family. It accepts mixed text + image input through the standard Chat Completions, Messages, and Responses APIs. On the text side it's on par with DeepSeek-V4-Flash; on visual agent benchmarks DeepSeek reports it landing close to Opus 4.8.
The numbers that matter
The pricing is identical to flash — $0.22 per million input tokens (cache miss), $0.66 per million output, off-peak. Images are converted to tokens based on their dimensions and billed as input tokens. That makes it essentially free to add vision to an agent you already run on flash.
Putting it to the test
I switched my agent's main model to it and threw the kind of input that usually breaks vision models: a dense analytics dashboard. Small monospace text, dollar figures, percentage deltas, a four-row transaction table, axis labels on a bar chart.
It read everything correctly. Not one number off. The customer names, the transaction IDs, the amounts, the statuses, the percentages — all exact.
The trade-off: it's experimental
The "Exp" suffix means this is an experimental release. It can change or have rough edges while it stabilizes. So I kept a pragmatic split:
- Main model (chat):
deepseek-v4-flash-vision-exp - Delegation + cron (background):
deepseek-v4-flash— the stable fallback - Auxiliary vision:
gemini-3-flash— a second fallback for tool channels
That way, if the experimental model misbehaves in chat, the long-running background work doesn't silently break.
Why this matters
For anyone building agents, the interesting trend isn't the top-of-the-line frontier models. It's that the cheap tier keeps absorbing premium capabilities. Vision was once a premium feature; now it's bundled into the $0.22/M model you probably already use.
Worth watching where this goes.
Source: DeepSeek API changelog, August 21, 2026.
