DEEPSEEK / MODEL PROFILE
DeepSeekReasoning with an eye on cost.
V4 Pro · V4 Flash
A useful comparison for high-volume text work, coding, and teams evaluating downloadable model weights.
Research checked September 5, 2026
THE ASSESSMENT
Documented capabilities and limitations, with practical interpretation.
Tokens are small units used to represent inputs and outputs. API rates are usage charges for software, separate from chat subscriptions. Prices are in US dollars, as checked for this edition.
Read our research method →Where it stands out
Flexible reasoning for repeated work
The V4 API offers thinking and non-thinking modes, a million-token context window, and tool calls. The August Pro update adds selectable reasoning effort and native Responses API support for agent integrations.
DeepSeek · V4 Pro general availability ↗DeepSeek · Current models and pricing ↗A lower-cost API comparison
V4 Pro’s published uncached input/output rates are $0.66 / $1.98 per million tokens off-peak and $1.32 / $3.96 at peak. Flash is cheaper. Scheduling can therefore matter as much as choosing a tier.
DeepSeek · Current models and pricing ↗Downloadable weights
The V4 Pro repository provides model weights under an MIT license. That gives a technically equipped team a route to control its deployment, rather than relying exclusively on the provider’s hosted application.
DeepSeek · Open model card and license ↗Where to be careful
Vision is a separate experiment
The current catalog separates V4 Pro and Flash from V4 Flash Vision Exp. Do not assume a scanned document or image will work with the ordinary text endpoint merely because another V4 variant accepts it.
DeepSeek · Current models and pricing ↗Open weights are not a desktop appliance
The Pro card describes a 1.6-trillion-parameter model, with 49 billion active parameters. Self-hosting the full model is substantial infrastructure work. You also become responsible for access controls, monitoring, and updates.
DeepSeek · Open model card and license ↗Version and price comparisons can mislead
The API has changed since the preview benchmarks, and peak/off-peak billing replaced the original prices in August. Evaluate the exact endpoint you plan to use, rather than applying an older V4 result or price to it.
DeepSeek · V4 Pro general availability ↗DeepSeek · Open model card and license ↗OUR SUGGESTED TRIAL
Try it on your kind of work.
Run a batch of fictional support messages through Pro and Flash. Ask for a category, a short rationale, and a draft response in a fixed format. Include ambiguous examples that should be escalated to a person.
What to look for
Measure correct classifications, unnecessary escalations, invented promises, total cost, and human correction time. Repeat the same examples at more than one reasoning setting.
An editorial test idea, not a reported benchmark result. Use public, fictional, or properly approved material.What else to compare
Start a cost comparison with V4 Flash, then test whether Pro improves your difficult examples enough to justify the difference. Compare a hosted service and a self-managed deployment as separate systems.
DeepSeek · Current models and pricing ↗No model in this guide replaces source checking, professional judgment, or a review of the application’s data terms.
Sources & further reading
Primary documentation checked for this edition. Provider claims are not an independent head-to-head test.