DeepSeek V4 Guide: Models, Access, Pricing and Limits
A source-backed guide to DeepSeek V4-Pro and V4-Flash using DeepSeek's release notes, pricing page, and transparency center.
What DeepSeek released
DeepSeek announced the V4 preview on April 24, 2026. The API exposes two model IDs: deepseek-v4-pro and deepseek-v4-flash. Both support thinking and non-thinking modes through OpenAI-compatible and Anthropic-compatible interfaces.
DeepSeek publishes model cards and technical material through its transparency center. Its launch post describes V4-Pro as a 1.6 trillion parameter mixture-of-experts model with 49 billion active parameters, and V4-Flash as a 284 billion parameter model with 13 billion active parameters. Those figures are provider-published specifications.
Model choice
| Decision | V4-Pro | V4-Flash |
|---|---|---|
| Primary role | Higher-capability reasoning and agent work | Faster, lower-cost general and agent work |
| API ID | deepseek-v4-pro | deepseek-v4-flash |
| Published context | 1 million tokens | 1 million tokens |
| Thinking mode | Supported | Supported |
| Best first test | Complex coding and multi-step reasoning | High-volume workflows and routing baseline |
Access and pricing
The API base URL remains https://api.deepseek.com. DeepSeek also publishes an Anthropic-compatible base URL. Current prices are listed in Chinese yuan per million tokens and differ for cache hits, uncached input, and output. Because those rates can change, this guide links to the live pricing page rather than freezing a price claim into the recommendation.
The older deepseek-chat and deepseek-reasoner names were scheduled for retirement on July 24, 2026. New integrations should use the V4 model IDs and confirm the current migration notes.
Capabilities and evidence limits
DeepSeek presents V4 as an agentic, coding, reasoning, and long-context model family. It also publishes open weights and a technical report. Benchmark statements in the release post remain provider measurements; they should not be treated as an independent result or as proof that V4 will outperform a closed model on every workload.
Practical implementation checklist
- Use the current V4 model IDs instead of retired aliases.
- Set thinking mode and effort deliberately rather than accepting defaults without testing.
- Measure cache-hit behaviour because it can materially change cost.
- Apply user isolation, rate limits, logging, and permission boundaries to agent workflows.
- Evaluate both Pro and Flash on the same task set before choosing a default.
Hussein's Take
V4-Flash should be the first cost and latency baseline for most teams; V4-Pro should earn its place on tasks where additional reasoning changes the result. DeepSeek's open technical material is useful, but production selection still requires your own task-level evaluation.