The 0813 build ends a preview that ran since April, pinning DeepSeek's top tier to a stable version behind the deepseek-v4-pro endpoint — a 1.6-trillion-parameter, MIT-licensed model with a 1M-token context and API rates far below Western frontier labs.
2026-08-12 · by Moe Ameen
On August 12, 2026, DeepSeek shipped DeepSeek-V4-Pro-0813, the general-availability build of its flagship V4 Pro model. It ends a preview that had run since April 24, 2026, when DeepSeek first released the V4 series with open weights, and it follows the smaller V4-Flash tier, which reached official status on July 31, 2026. The 0813 build now sits behind DeepSeek's `deepseek-v4-pro` API endpoint as a stable, versioned model ID, and its weights remain open under the MIT license.
V4 Pro is a mixture-of-experts model with roughly 1.6 trillion total parameters and about 49 billion active per token. It carries a context window near 1,048,576 tokens and a maximum output around 384,000 tokens, and DeepSeek describes an attention design — a compressed sparse variant plus a heavily compressed one — that the company says cuts single-token inference compute to about 27% and KV cache to roughly 10% of what its V3.2 generation needed at the million-token setting. The API exposes three operating modes: a fast non-thinking mode, a high-reasoning-effort mode, and a max-effort mode.
On the GA build DeepSeek reports strong benchmark figures, weighted toward coding and agentic work: about 80.6% on SWE-bench Verified, 93.5% pass@1 on LiveCodeBench, 90.1% on GPQA Diamond, and a Codeforces rating near 3,206. Pricing carries over from the preview — roughly $0.435 per million input tokens on a cache miss, about $0.0036 per million on a cache hit, and $0.87 per million output — which keeps it well below comparable frontier models from OpenAI, Anthropic, and Google. DeepSeek has not detailed every training change from the preview weights; treat the specific figures as launch-day numbers and confirm them on deepseek.com before relying on them. Like the rest of the family, 0813 is a text-and-reasoning model: it writes, reasons, and codes, but it generates no images, video, or audio and publishes nothing.
The fastest move on this release is not to open the DeepSeek API — it is to publish your read on it. "The cheapest frontier flagship just hit GA" is a beat your audience is already seeing, so being early with a sharp take beats being thorough a week late. Feed your angle into [Kompozy](/) as a source and it becomes a [Blog Article](/glossary/output-buckets) on what a cheap open-weight flagship changes for content teams, a [Carousel](/glossary/output-buckets) laying out where 0813 sits against closed frontier models, a few captioned shorts, and platform-native posts in your voice through the [Persona Brief](/glossary/persona-brief) — scheduled across the eight social platforms plus blog and email from one queue with [Autopilot](/glossary/autopilot).
The second move is for when you actually draft with it. 0813 is an outstanding upstream brain, and because Kompozy treats every model and generator as an interchangeable input, the fact that top-tier drafting just got cheaper is pure upside — swap it in as your scripting layer without rebuilding anything downstream. Draft a whole week of scripts and outlines in one long 0813 call, then hand each to Kompozy to render the media the model can't: [Persona Shorts](/glossary/persona-shorts) avatar video, brand-exact carousels, [Quote Graphics](/glossary/output-buckets) of the sharpest lines, a formatted blog, and a newsletter — all held to one look by [HyperFrames](/glossary/hyperframes) and published across nine destinations. The model keeps getting cheaper and stronger; the part that still decides whether anyone sees the work is the captioning, formatting, and distribution Kompozy owns. One note: Kompozy's own copy generation runs on Claude and OpenAI, so 0813 is the drafting choice feeding the pipeline, not a swap for it.
DeepSeek V4 Pro 0813 is the general-availability build of DeepSeek's flagship V4 Pro model, released on August 12, 2026 after a preview that began April 24, 2026. It is a mixture-of-experts LLM (roughly 1.6 trillion total / 49 billion active parameters) with about a 1-million-token context window and MIT-licensed open weights, served behind DeepSeek's deepseek-v4-pro endpoint.
On DeepSeek's first-party API it is roughly $0.435 per million input tokens on a cache miss, about $0.0036 per million on a cache hit, and $0.87 per million output — well below comparable Western frontier models. Because the weights are open under the MIT license, you can also self-host to avoid per-token fees. Confirm current rates on deepseek.com.
It can draft the text — scripts, captions, outlines — but no more. It is a text-and-reasoning model with no image, video, or audio generation and no publishing. Turning its drafts into finished carousels, avatar video, and scheduled multi-platform posts is a separate job handled by a content engine like Kompozy.
Draft a batch of scripts and outlines in DeepSeek V4 Pro 0813, then bring them into Kompozy to render Persona Shorts, carousels, quote graphics, a blog, and a newsletter in your brand voice, and schedule and publish across the eight social platforms plus blog and email from one queue.