--- title: DeepSeek's V4 Pro release and what it means for OpenAI's... description: DeepSeek released V4 Pro (build 0813), a post-training upgrade built from distilled specialist checkpoints for math, coding, and agentic tasks, claiming up to... canonical: https://daily.dev/posts/deepseek-s-v4-pro-release-and-what-it-means-for-openai-s-timing-btiz9gqb9 twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: DeepSeek's V4 Pro release and what it means for OpenAI's timing | daily.dev og:description: DeepSeek released V4 Pro (build 0813), a post-training upgrade built from distilled specialist checkpoints for math, coding, and agentic tasks, claiming up to... og:url: https://daily.dev/posts/deepseek-s-v4-pro-release-and-what-it-means-for-openai-s-timing-btiz9gqb9 og:image: https://api.daily.dev/og/posts/BtIz9GQB9.png og:image:alt: DeepSeek's V4 Pro release and what it means for OpenAI's timing og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # DeepSeek's V4 Pro release and what it means for OpenAI's timing **[Collections](https://daily.dev/sources/collections)** · 3 min read · 1 upvotes · 0 comments ## Summary DeepSeek released V4 Pro (build 0813), a post-training upgrade built from distilled specialist checkpoints for math, coding, and agentic tasks, claiming up to 78% faster generation via improved multi-token drafting. Weights are MIT-licensed and open, though DeepSeek's own hosted API price jumped 2.5x-5x. A companion open-source coding harness (Cordis) was tested against Claude Code and Codex, holding up surprisingly well though not quite matching Claude Code's polish. The piece speculates that OpenAI's announced two-week pause on frontier RL training before its next model (codenamed Astra) may be a reaction to DeepSeek's release rather than a planned safety checkpoint. ## Content ## What was released DeepSeek released V4 Pro (build 0813), a 320B-parameter mixture-of-experts model with 18B active parameters, under the MIT license. That means anyone can self-host it or run it through third-party providers. DeepSeek's own hosted API did get more expensive - prices jumped roughly 2.5x to 5x over the earlier preview - but the open weights mean you're not locked in. ## What actually changed under the hood This isn't a new architecture. The gains come from post-training: DeepSeek trained multiple specialist checkpoints (math, coding, agentic tasks) and distilled them into a single final model. They also improved multi-token speculative decoding, which they claim yields up to 78% faster generation. The distillation and speculative-decoding technique came from a research paper published roughly six weeks before the model dropped, which is a fast turnaround from paper to production. Compared to the earlier preview and flash versions, V4 Pro shows noticeably better spatial and structural reasoning. Early testing with one-shot app-building prompts puts it in the same conversation as Claude Code and Codex, though hands-on comparisons vary depending on the task. ## The repository-level engineering angle The more interesting claim is around agentic, repository-level work. Generating a function from a clear prompt is one thing. Real codebases have multiple files, modules, dependencies, existing conventions, and tests where changing one component can break something three layers away. DeepSeek says V4 Pro is specifically improved for this kind of work, with gains reported on DeepSWE, CyberGym, and NL2Repo benchmarks. In practice, that means the model can: - Read a project's overall structure and locate relevant files - Plan changes across multiple components while preserving existing behavior - Run tests, analyze failures, and revise without losing context - Handle tasks like multi-file feature implementation, refactoring, dependency updates, and security review That said, human judgment still matters here. Someone needs to define permissions, review proposed changes, validate tests, and approve anything going to production. V4 Pro expands how deeply the model can participate in engineering work - it doesn't replace the engineer making the calls. ## The broader context OpenAI announced a two-week pause on frontier reinforcement learning training ahead of a new model, which some observers read as a competitive response to pressure from Chinese labs. Whether or not that's the right read, DeepSeek releasing a capable open-weight model under MIT while the closed providers charge more for access is a real dynamic worth watching. The gap between what you can run yourself and what requires a proprietary API keeps narrowing. ## Questions this post answers ### What are the specs of DeepSeek V4 Pro build 0813? DeepSeek V4 Pro (build 0813) is a 320B-parameter mixture-of-experts model with 18B active parameters, released under the MIT license so it can be self-hosted or run through third-party providers. DeepSeek's own hosted API pricing rose roughly 2.5x to 5x compared to the earlier preview version. _Engineers weighing open-weight versus proprietary coding models can track releases like this one on daily.dev._ ### How much faster is DeepSeek V4 Pro's generation compared to previous versions? DeepSeek claims up to 78% faster generation in V4 Pro, achieved through improved multi-token speculative decoding rather than a new architecture. The gains also come from distilling multiple specialist checkpoints trained separately for math, coding, and agentic tasks into a single final model. _Anyone benchmarking inference speed across model releases can follow updates like this via daily.dev._ ### What benchmarks show DeepSeek V4 Pro's repository-level coding improvements? DeepSeek reports gains on DeepSWE, CyberGym, and NL2Repo benchmarks, which test repository-level engineering work such as locating relevant files, planning multi-file changes, running tests, and revising code without losing context. Human review of permissions, changes, and test validation is still required before anything reaches production. _Developers evaluating agentic coding tools for real codebases can compare capabilities like these on daily.dev._ --- Tags: [#machine-learning](https://daily.dev/tags/machine-learning), [#llm](https://daily.dev/tags/llm), [#openai](https://daily.dev/tags/openai), [#ai-coding](https://daily.dev/tags/ai-coding), [#deepseek](https://daily.dev/tags/deepseek) [View this post on daily.dev](https://daily.dev/posts/deepseek-s-v4-pro-release-and-what-it-means-for-openai-s-timing-btiz9gqb9) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"DeepSeek's V4 Pro release and what it means for OpenAI's timing","url":"https://daily.dev/posts/deepseek-s-v4-pro-release-and-what-it-means-for-openai-s-timing-btiz9gqb9","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/deepseek-s-v4-pro-release-and-what-it-means-for-openai-s-timing-btiz9gqb9"},"datePublished":"2026-08-20T17:57:29.629Z","dateModified":"2026-09-13T20:02:42.440Z","description":"DeepSeek released V4 Pro (build 0813), a post-training upgrade built from distilled specialist checkpoints for math, coding, and agentic tasks, claiming up to...","image":"https://i.ytimg.com/vi/xBByvFrqmWU/sddefault.jpg","thumbnailUrl":"https://i.ytimg.com/vi/xBByvFrqmWU/sddefault.jpg","isAccessibleForFree":true,"articleSection":"Collections","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Collections","logo":"https://media.daily.dev/image/upload/s--fk_6ycEi--/f_auto,q_auto/v1780996001/logos/collections?_a=BAMAMiWQ0","url":"https://daily.dev/sources/collections"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/deepseek-s-v4-pro-release-and-what-it-means-for-openai-s-timing-btiz9gqb9","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":1},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"machine-learning,llm,openai,ai-coding,deepseek","timeRequired":"PT3M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Collections","item":"https://daily.dev/sources/collections"},{"@type":"ListItem","position":3,"name":"DeepSeek's V4 Pro release and what it means for OpenAI's timing"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/deepseek-s-v4-pro-release-and-what-it-means-for-openai-s-timing-btiz9gqb9#faq","mainEntity":[{"@type":"Question","name":"What are the specs of DeepSeek V4 Pro build 0813?","acceptedAnswer":{"@type":"Answer","text":"DeepSeek V4 Pro (build 0813) is a 320B-parameter mixture-of-experts model with 18B active parameters, released under the MIT license so it can be self-hosted or run through third-party providers. DeepSeek's own hosted API pricing rose roughly 2.5x to 5x compared to the earlier preview version. Engineers weighing open-weight versus proprietary coding models can track releases like this one on daily.dev."}},{"@type":"Question","name":"How much faster is DeepSeek V4 Pro's generation compared to previous versions?","acceptedAnswer":{"@type":"Answer","text":"DeepSeek claims up to 78% faster generation in V4 Pro, achieved through improved multi-token speculative decoding rather than a new architecture. The gains also come from distilling multiple specialist checkpoints trained separately for math, coding, and agentic tasks into a single final model. Anyone benchmarking inference speed across model releases can follow updates like this via daily.dev."}},{"@type":"Question","name":"What benchmarks show DeepSeek V4 Pro's repository-level coding improvements?","acceptedAnswer":{"@type":"Answer","text":"DeepSeek reports gains on DeepSWE, CyberGym, and NL2Repo benchmarks, which test repository-level engineering work such as locating relevant files, planning multi-file changes, running tests, and revising code without losing context. Human review of permissions, changes, and test validation is still required before anything reaches production. Developers evaluating agentic coding tools for real codebases can compare capabilities like these on daily.dev."}}]} ```