DeepSeek V4 Pro 0813 Goes GA: 1M-Context, Agent-Focused, Under a Dollar per Million Tokens

DeepSeek V4 Pro 0813 left preview on August 13 — 1.6T-parameter MoE with 1M-token context, agentic tool use, and sub-dollar pricing before a price hike lands.

DeepSeek's V4 Pro (build 0813) reached general availability on August 13, becoming the lab's new flagship. The architecture is a mixture-of-experts system with 1.6 trillion total parameters and roughly 49 billion active per forward pass. The 0813 release adds DSpark speculative decoding for faster output and exposes configurable reasoning-effort levels (low / high / max) per request.

Key specs:

• Context window: 1 million tokens (384K max output)
• Pricing: $0.435/1M input (cache miss) · $0.003625/1M cached · $0.87/1M output
• Native OpenAI Responses API support (one-click Codex-style setup)
• Reasoning effort: low, high, or max per request — same model, variable cost

The model is positioned around agentic workloads: multi-step tool use, long-horizon reasoning, and code generation. Vendor-reported SWE-bench scores are striking (up to ~91%), but no independent evaluator has replicated them as of publication.

Why it matters: At $0.435/1M input (and sub-cent cached), V4 Pro undercuts most closed frontier models for high-context agentic pipelines. The 1M-token window means entire codebases or document collections can fit in a single call.

Heads up: DeepSeek has announced a price increase is coming — no timeline, no figure, no affected tiers disclosed. Current pricing likely reflects a promotional discount. Factor in a conservative multiplier before baking these numbers into production cost projections.

Available via the DeepSeek API and through OpenRouter.

Sources: DeepSeek Ships V4 Pro — Unite.AI · DeepSeek V4 Pro Review — CoderSera