DeepSeek's vision model undercuts Gemini on price, but can't match its speed

DeepSeek unveiled its first vision-enabled model, V4 Flash Vision Exp, on August 21, marking the company's entry into image recognition alongside text processing. The new capability enables the system to parse charts, screenshots, and document images with the same proficiency it applies to written input.

Now accessible through API platforms like OpenRouter, DeepSeek's vision model maintains the parent V4 Flash's ultra-low pricing at $0.22 per million input tokens and $0.66 per million output tokens—though costs double during weekday peak hours. Google's Gemini 3.7 Flash, released two weeks earlier, stands as the reference point for budget-conscious developers, priced at $0.75 and $3.75 per million tokens on the same platform.

How the models were tested#

To determine which excels in real-world scenarios, three practical exercises mimicked common back-office workloads:

• Chart reading: A stacked bar chart with dual y-axes at different scales, designed to trip up models that eyeball rather than calculate
• Invoice audit: A vendor document containing three intentional errors—a miscalculated line item, a subtotal mismatch, and a due date before the invoice date
• Log analysis: Production logs featuring a payment service failure, with the actual cause (a database connection pool exhaustion from a batch job) appearing earlier in the trace

Accuracy: A clean sweep for both#

Neither model stumbled on the planted traps. On the dual-axis chart, both correctly identified Q1 as the period when costs exceeded revenue, named Subscriptions as the only segment growing all four quarters, and calculated annual revenue at $36.1 million—matching the source data exactly.

For the invoice audit, each model caught all three errors, pinpointed the miscalculated line ($505.89 instead of $459.90), and recalculated the correct total as $3,958.89. Both flagged the chronologically impossible due date. DeepSeek went further, spotting that the printed subtotal didn't align with individual line items.

On the production logs, both correctly bypassed the payment service crash decoy and identified the batch analytics job as the root cause, pinpointing 14:05:12 as when the problem began. Gemini even supplied specific PostgreSQL commands for remediation.

Cost and speed: Where they diverge#

Performance metrics revealed the crucial tradeoff. Gemini answered in an average of 7.2 seconds across all tests, while DeepSeek took 16.8 seconds—more than double. The invoice test showed the most dramatic difference: DeepSeek consumed 30.5 seconds and 3,467 completion tokens for a brief response, suggesting substantial hidden reasoning overhead, versus Gemini's 7.9 seconds and 944 tokens.