DeepSeek's new V4 Flash Vision Exp matched Google's Gemini 3.7 Flash in accuracy across three real-world image understanding tests while charging roughly one-third the price, according to a head-to-head comparison published by The New Stack. Released August 21, the model represents DeepSeek's first entry into vision-capable AI, adding image comprehension to its budget Flash model while maintaining pricing of $0.22 per million input tokens and $0.66 per million output tokens. The tests found identical performance on document analysis, chart interpretation, and log diagnosis, with differences emerging only in speed and cost.
Both systems answered every question correctly across three simulated back-office tasks designed to trap weaker models. In a chart reading exercise featuring dual y-axes with different scales, both correctly identified Q1 as the quarter when operating costs surpassed total revenue, named Subscriptions as the only segment that expanded every quarter, and calculated full-year revenue at exactly $36.1 million. An invoice audit containing three planted errors—a line total that didn't match quantity times price, a subtotal unrelated to any line items, and a due date preceding the invoice date—caught neither model off guard, with each flagging all three mistakes and arriving at the correct total due of $3,958.89. In a production log analysis where a payment service crash at the bottom concealed the true culprit five minutes earlier, both ignored the decoy and pinpointed a manually triggered analytics job consuming all 20 database connections at 14:05:12 as the root cause.
The key divergence appeared in processing speed and token consumption, the tester reported. Gemini responded in an average of 7.2 seconds compared to DeepSeek's 16.8 seconds, with DeepSeek's timing fluctuating from 8 to 30 seconds depending on the task. DeepSeek's total bill reached $0.0039 against Gemini's $0.0122, though DeepSeek doubles its rates during weekday peak hours and testing occurred on a weekend. The invoice test produced an outlier: DeepSeek required 30.5 seconds and billed 3,467 completion tokens for a response only a few paragraphs long, suggesting substantial internal reasoning charged as output, while Gemini finished in 7.9 seconds with 944 completion tokens.
The practical implications depend on scale and use case, the tester concluded. For individual users processing small volumes, the difference between 7 and 17 seconds carries little weight, and neither price point accumulates quickly at low volume. DeepSeek suits batch processing scenarios like overnight invoice reconciliation where users aren't waiting, while Gemini fits real-time applications where someone expects an immediate answer. The two companies tokenize images in fundamentally different ways, with DeepSeek counting each image at roughly 500 prompt tokens versus Gemini's approximately 1,150, indicating divergent technical approaches beneath identical accuracy. Organizations choosing between vision models now face a straightforward trade-off rather than a quality gap: pay less and wait longer, or pay more for instant results. The appearance of cost-competitive alternatives in the vision AI space challenges the assumption that superior speed or brand recognition justify premium pricing when accuracy remains indistinguishable.