DeepSeek V4-Pro Goes Live and Runs OpenAI's Own Coding Agent 8x Cheaper
DeepSeek-V4-Pro exits four months of preview with major agent upgrades, flexible reasoning modes, and native OpenAI Responses API support for Codex workflows
- GA release: DeepSeek-V4-Pro exits preview as build
V4-Pro-0813, available on web, mobile, and API now. - Agent upgrades: Major post-training improvements for agentic workflows; flexible reasoning effort (low/high/max) across V4-Pro and V4-Flash.
- Codex support: Native OpenAI Responses API integration lets you run Codex on DeepSeek with one-click setup -- no bridging code needed.
- Pricing: V4-Pro at $0.435/M input and $0.87/M output; peak/off-peak rates (50% off-peak discount) take effect August 16, 2026.
- Benchmarks: 80.6% on SWE-bench Verified (near Claude Opus 4.6), 93.5% LiveCodeBench -- at ~1/7 the output price of frontier competitors.
- Architecture: 1.6T total / 49B active MoE, 1M-token context, 384K max output; MIT-licensed open weights (April preview build) on Hugging Face.
DeepSeek-V4-Pro has officially left preview. After nearly four months of staged rollout, the flagship model is now fully available on the web app, mobile app, and API under the build designation V4-Pro-0813, completing a two-part release strategy that started when DeepSeek shipped V4-Flash to general availability on July 31.
What's Under the Hood
V4-Pro is a mixture-of-experts system with 1.6 trillion total parameters and 49 billion active per token. Mixture-of-experts means the model is divided into many specialized sub-networks, and only a small fraction of them activate for any given token, so you get the capacity of a 1.6T model at the compute cost of a much smaller one. Both V4-Pro and V4-Flash share a 1M-token context window.
The attention system combines Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA), targeting serving cost rather than raw capability. These are DeepSeek's custom attention variants that dramatically reduce the memory required to process long contexts. The KV cache (the memory buffer that stores past tokens during inference) drops to 10% of what V3.2 needed at the million-token setting. V4-Flash totals 284 billion parameters with around 13 billion active per token.
What Actually Changed in the GA Build
The architecture is the same one that shipped in April. Post-training is where the work went. When V4-Flash went official, DeepSeek published agent-benchmark results showing the re-post-trained Flash build outscoring the V4-Pro-Preview on its internal coding-agent suites, a deliberate move that made the small model the default for agent workloads while the flagship stayed in preview. The 0813 build is the flagship's answer to that.