GLM-52 897 —
GPT-56SC 873 —
CL-OP5X 865 -0.9%
GROK-46H 865 -0.9%
GEM-37FH 865 -0.9%
GPT-56T 861 —
GLM-5 856 —
MUSE-SPK 841 —
QWEN-38X 824 -2.3%
GPT-6A 820 —
KIMI-K3X 810 -1%
CL-FAB5H 787 -0.9%
CL-OP5H 764 -0.9%
CL-OP46H 742 -0.9%
CL-OP47H 733 -1.1%
GEM-38FH 676 -1%
CL-OP47 585 -0.7%
INKL 531 —
CL-OP46 496 -0.2%
CL-OP48 490 -0.2%
GLM-52 897 —
GPT-56SC 873 —
CL-OP5X 865 -0.9%
GROK-46H 865 -0.9%
GEM-37FH 865 -0.9%
GPT-56T 861 —
GLM-5 856 —
MUSE-SPK 841 —
QWEN-38X 824 -2.3%
GPT-6A 820 —
KIMI-K3X 810 -1%
CL-FAB5H 787 -0.9%
CL-OP5H 764 -0.9%
CL-OP46H 742 -0.9%
CL-OP47H 733 -1.1%
GEM-38FH 676 -1%
CL-OP47 585 -0.7%
INKL 531 —
CL-OP46 496 -0.2%
CL-OP48 490 -0.2%
← Back to feed

DeepSeek V4 Pro 0813 Goes GA After Four-Month Preview — 1.6T MoE Becomes the Official Flagship

DeepSeek promoted V4 Pro to general availability on August 12, 2026, designating the 0813 build as the production version of its flagship model. The API documentation now resolves deepseek-v4-pro to DeepSeek-V4-Pro-0813, ending a preview period that began when DeepSeek shipped open weights and API access for both V4 Pro and V4 Flash on April 24, 2026.

The Model

V4 Pro is a mixture-of-experts architecture with 1.6 trillion total parameters and 49 billion active per token. Two attention variants — Compressed Sparse Attention and Heavily Compressed Attention — reduce single-token inference compute to 27% and KV cache to 10% of what V3.2 required at the million-token context setting.

Pre-training ran on over 32 trillion tokens. Post-training grew domain-specific experts separately and consolidated them through on-policy distillation.

SWE-bench Verified: 80.6%, established at the April preview release and unchanged through GA.

Pricing and Limits

TierRate
Input (cache miss)$0.435 / M tokens
Input (cache hit)$0.003625 / M tokens
Output$0.87 / M tokens
Context1M tokens
Max output384K tokens
Concurrency500

V4 Flash holds a higher concurrency ceiling at 2,500, positioning it for bulk inference workloads where throughput outweighs capability.

What Changed With GA

The model weights are unchanged. Open weights for the 0813 build have been available on Hugging Face under MIT license since April. The GA designation updates the API contract and provides a stable versioned endpoint for production deployments that need to pin to a specific build.

On July 31, DeepSeek promoted V4 Flash to GA and stated the Pro release would follow. The 0813 build fulfils that commitment on the 12-day timeline it implied.

Competitive Context

At $0.87 per million output tokens, V4 Pro undercuts GPT-5.6 Sol ($30) and Claude Fable 5 ($75) by an order of magnitude. DeepSeek V4 Flash sits even lower at $0.14/M output. The frontier pricing gap between Chinese and US labs has not closed — if anything, the V4 GA makes it more visible as a stable, production-grade endpoint rather than a rolling preview.