DeepSeek V4 Pro 0813

BenchLM's DeepSeek provider overview confirms that the deepseek-v4-pro API ID currently serves DeepSeek V4 Pro 0813, a 1M-token context model that supports both thinking and non-thinking modes, and notes the model's retirement status as DeepSeek phases it out in favor of V4.1 Flash. The page provides model-family-level The same overview indicates DeepSeek is not recommending new routing rules around V4 Pro 0813 given V4.1 Flash's reported performance and pricing advantages, and that legacy deepseek-v4-flash requests were redirected to V4.1 Flash after September 10, 2026. This makes the candidate relevant for readers tracking V4 Pro 0

DeepSeek-V4-Pro-0813 shipped generally available on August 13, 2026, across DeepSeek's app, web interface, and API with a quiet announcement that emphasised "significantly enhanced agent capabilities" before that language was reportedly pulled from the company's site shortly afterwards. The release introduces a hybrid The hybrid attention design is paired with new techniques for stabilising very deep residual stacks, and the article references an underlying research paper describing reduced compute and memory costs at the million-token setting compared with the previous generation. Independent benchmarking organisations reportedly f

China's National Supercomputing Internet announced on August 15, 2026 that the official version of DeepSeek-V4-Pro-0813 and the DeepSeek Harness agent framework are live on its platform, backed by what the platform describes as China's first 100,000-accelerator resource pool integrating supercomputing and AI computing. Alongside the 0813 release, DeepSeek Harness — an MIT-licensed agent framework with an "everything is a plugin" architecture — was open-sourced, supporting freely replaceable models, tools, skills and sessions, and offering four operating modes: Standard, PTC, Minimalist, and Creative. A Chinese expert quoted in the pi

MorphLLM's V4 guide explicitly names the V4-Pro-0813 checkpoint and details its architecture: a 1.6-trillion-parameter MoE with 49B active parameters, a 1-million-token context window, 384K maximum output, MIT license, and a GA date of August 13, 2026. Off-peak API pricing is listed at $0.66 input and $1.98 output per The page also documents self-hosting references including an 8x B200 node with SGLang/NVFP4 or 4x GB300 with vLLM, citing Lambda pricing at $53.52/hr, and lists first-party deployment options such as api.deepseek.com, OpenRouter, Cloudflare Workers AI, DeepInfra, Together, Lightning, NVIDIA NIM, and LM Studio for local

An August 18, 2026 review confirms DeepSeek V4 Pro 0813 as a 1.6-trillion-parameter MoE model with 49B active parameters, 1M-token context window, and MIT-licensed open weights. DeepSeek reports 87.9 on Terminal-Bench 2.1, but independent measurement by CoderSera on a neutral harness yielded 54.68%, while V4 Flash reac Current official pricing is $0.66/$1.32 per million cache-miss input tokens and $1.98/$3.96 per million output tokens (off-peak/peak), with cheaper cached input rates. The piece frames 0813 as a production build focused on serving efficiency, adding roughly 51.7 billion parameters across four new speculative-decoding "

DeepSeek moved DeepSeek-V4-Pro-0813 to general availability on August 13, 2026, ending the preview period that began with the V4 series in April 2026. The deepseek-v4-pro endpoint now serves the 0813 build without changing the callable model name, and the model retains its one million token context window. DeepSeek als On DeepSeek's own evaluation harness run in minimal mode, DeepSeek-V4-Pro-0813 scores 87.9 on Terminal-Bench 2.1, compared with 72.1 for the April V4-Pro preview, and the release also adds a selectable reasoning effort with low, high, and max levels (replacing the earlier fixed thinking budget) alongside native support

MindStudio published a benchmark analysis titled "DeepSeek-V4-Pro-0813 Benchmarks: How It Stacks Up Against Opus and Kimi K3," explicitly naming the 0813 variant. The blog post is a third-party comparative analysis of the model's performance against competing frontier models, providing external perspective on where Dee The scraped excerpt is dominated by cookie consent banner text, which limits the extraction of detailed benchmark numbers and analysis from the blog post itself. The article is positioned as a technical comparison piece for developers evaluating the model against alternatives, though specific benchmark figures and deta

Remio's August 14, 2026 analysis confirms the August 13, 2026 general-availability rollout of DeepSeek-V4-Pro-0813 across DeepSeek's app, web service, and API, with existing API users able to access it via the deepseek-v4-pro model name. The post highlights tension between DeepSeek's internal benchmarks, which show str The article grounds its technical claims in the official V4 model card, citing the mixture-of-experts architecture at 1.6 trillion total parameters with 49 billion active per token and a 1-million-token context window. It frames the 0813 release as a move from April preview to broader production with a coding-agent foc

DeepSeek officially released DeepSeek-V4-Pro-0813 on August 13, 2026, making the GA version available across its app, web interface, and API after a preview since April. The model focuses on agent capabilities and scored 87.9 on Terminal Bench 2.1, 62.7 on DeepSWE, and 61.5 on NL2Repo among other agent-focused tests. I The V4-Pro API has been updated to work with the OpenAI Responses API format out of the box and includes built-in support for Codex integration, while thinking effort levels have been expanded to three settings — low, high, and max. A price increase for the V4 family takes effect at 16:00 UTC on August 16, with peak an

South China Morning Post reported on August 13, 2026 that Chinese AI start-up DeepSeek quietly released DeepSeek-V4-Pro-0813, an updated version of its flagship model. The article describes the release as having left some developers underwhelmed by its overall capabilities and disappointed in its pricing. SCMP's coverage highlights that the updated model impressed researchers in niche areas such as cybersecurity despite the broader mixed reception on general capabilities. This represents original independent press coverage naming the exact 0813 variant on its launch date.