---
type: analysis
slug: research-deepseek-v4-pro-0813-is-a-real-ga-checkpoint-live-on-deepseek-and-openrouter
title: "DeepSeek V4 Pro 0813 is a real GA checkpoint, live on DeepSeek and OpenRouter"
description: "DeepSeek shipped DeepSeek-V4-Pro-0813 as official V4 Pro on 13 August 2026, with MIT weights, a first-party API alias, and a live OpenRouter slug. Builders should treat OpenRouter as a multi-provider router, not the source of truth for identity, pricing, or features."
published_at: 2026-10-26T07:00:00.000Z
topics: ["multimodal","openai","tools","research"]
source_urls: ["https://api-docs.deepseek.com/updates/","https://api-docs.deepseek.com/news/news260813","https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro-0813","https://huggingface.co/collections/deepseek-ai/deepseek-v4","https://api-docs.deepseek.com/quick_start/pricing/","https://api-docs.deepseek.com/guides/thinking_mode/","https://openrouter.ai/deepseek/deepseek-v4-pro-0813","https://artificialanalysis.ai/models/deepseek-v4-pro","https://arxiv.org/abs/2606.19348"]
author: "Andrey Reshetnikov"
author_profile: https://newruntime.com/owner-profile.md
---
When OpenRouter listed deepseek/deepseek-v4-pro-0813, the newsroom treated that page as a distribution signal rather than a DeepSeek release. That caution was right as method, but the model is real. On 13 August 2026 DeepSeek posted a GA announcement, a changelog entry, and a Hugging Face card for DeepSeek-V4-Pro-0813. Its current Models & Pricing table still names that checkpoint as the model behind the first-party alias deepseek-v4-pro. OpenRouter’s model page, checked on 23 October 2026, describes the same GA release and routes it across many hosts, including a DeepSeek-branded provider.

This is the large text-only member of DeepSeek’s V4 family: MIT open weights, a one-million-token context, tool calling, native OpenAI Responses API support, and three thinking-effort levels. It superseded the April preview after DeepSeek said the Flash GA had already beaten Pro preview on several agent benches. For builders, the issue is no longer whether 0813 exists, but which endpoint, provider, price, and effort setting you are calling, and how that compares with the cheaper multimodal Flash line now behind deepseek-flash.

## What we checked

- **Did DeepSeek officially release DeepSeek-V4-Pro-0813 with a release note, model card, API announcement, and benchmarks?** confirmed. Yes. DeepSeek’s 13 August 2026 changelog and GA news post, the Hugging Face card deepseek-ai/DeepSeek-V4-Pro-0813, and the live Models & Pricing table all identify DeepSeek-V4-Pro-0813 as official V4 Pro, with published agent scores and API alias deepseek-v4-pro. [1][2][3][5]
- **Is OpenRouter currently listing and serving deepseek/deepseek-v4-pro-0813 as a live routed model?** confirmed. Yes. OpenRouter’s model page describes it as the GA V4 Pro, documents Chat Completions and Responses usage, lists many providers including a DeepSeek-branded row, and reported 100% three-day uptime and 99.71% availability in the window ending 23 October 2026. [7]

## What we found

- DeepSeek documented a GA rollout of V4 Pro on 13 August 2026 across app, web, and API. The changelog says the API method is unchanged: set model to deepseek-v4-pro. The same day’s news post lists agent upgrades, low/high/max reasoning effort, native OpenAI Responses API support adapted for Codex, and peak/off-peak pricing from 16:00 UTC on 16 August 2026. Official agent scores include HLE 42.7 without tools and 60.0 with tools, and Terminal Bench 2.1 at 87.9. [1][2]
- The official Hugging Face card states DeepSeek-V4-Pro-0813 is the official V4 Pro, superseding the preview, built on the preview structure with a DSpark speculative-decoding module attached. The same official agent numbers appear there. Weights are MIT-licensed. The card does not ship a Jinja chat template; encoding scripts live in an encoding folder. Hugging Face lists the safetensors size as 1.7T parameters, and the DeepSeek-V4 collection marks the 0813 repo as updated 13 August 2026. [3][4]
- DeepSeek’s English Models & Pricing page still maps API name deepseek-v4-pro to version DeepSeek-V4-Pro-0813. Context is 1M tokens and max output is 384K. Features include JSON output, tool calls, Responses API, and Anthropic API; vision is not supported. Peak prices are $0.044 per 1M cached input tokens, $1.32 per 1M cache-miss input, and $3.96 per 1M output; off-peak rates are half. Peak hours are 01:00–04:00 and 06:00–10:00 UTC, Monday through Friday, excluding Chinese public holidays. Concurrency limit is 500, versus 2500 for Flash. Thinking is on by default at high effort. [5][6]
- OpenRouter currently serves slug deepseek/deepseek-v4-pro-0813 as the GA V4 Pro, with a 1,048,576-token context and up to 393,216 completion tokens. The DeepSeek-branded provider row is $1.32 input and $3.96 output per 1M tokens, matching official peak list prices. The page’s headline in/out display was $0.2565 / $8 per 1M, which reflects a cheap-input, high-output host rather than DeepSeek’s tariff. Three-day uptime was 100.00% and availability 99.71% for 20–23 October 2026. The model is not available through EU in-region routing. [7]
- Artificial Analysis independently scores DeepSeek V4 Pro 0813 (Max) at 36 on its Intelligence Index, lists $1.32 / $3.96 first-party pricing, 1.6T total and 49B active parameters, a 1M context, text-only I/O, MIT license, and a 13 August 2026 release date. Its Terminal-Bench 2.1 score of 78.7% and HLE of 41.0% sit below DeepSeek’s published 87.9 and 42.7, which DeepSeek says used DeepSeek Harness minimal mode at max effort with temperature 1.0 and top_p 0.95. [8][3]
- On 10 September 2026 DeepSeek said it would keep V4 Pro API service after 14 September 2026 with billing unchanged. The Flash alias has since moved to DeepSeek-V4.1-Flash-1017, which is multimodal and cheaper. The April 2026 V4 paper describes the preview architecture—hybrid Compressed Sparse Attention and Heavily Compressed Attention, Manifold-Constrained Hyper-Connections, and 27% of DeepSeek-V3.2 single-token FLOPs and 10% of KV cache at 1M context for Pro. The 0813 card says it keeps that preview structure plus DSpark. [1][5][9][3]

## What it means

Treat 0813 as a production DeepSeek checkpoint, not an OpenRouter-only nickname. On DeepSeek’s API the string is deepseek-v4-pro; on OpenRouter it is deepseek/deepseek-v4-pro-0813. Pin a provider if you need first-party behavior or a known price, because routed hosts advertise very different input and output rates. Use first-party off-peak and cache-hit discounts if you can schedule work. Default thinking effort is high; on DeepSeek’s API, xhigh maps to high, not max. For tool loops, pass reasoning_content back or the first-party API returns 400. Do not send images to Pro.

If you mainly need cheaper, multimodal, or higher-concurrency serving, DeepSeek’s current Flash line is a different product and is no longer the 0731 V4 Flash checkpoint. Self-host from the MIT 0813 weights if you need a fixed binary; Hugging Face recipes enable DSpark in vLLM and SGLang from the same checkpoint. Do not mix OpenRouter’s 393,216 completion cap or third-party prices with DeepSeek’s 384K max-output and peak/off-peak tariff when writing evals or cost models.

## Still open

- Whether third-party OpenRouter hosts are serving the same 0813 weights and decoding stack as first-party DeepSeek, or only an API-compatible proxy.
- Why OpenRouter dates the release 12 August 2026 while DeepSeek, Hugging Face, and Artificial Analysis use 13 August 2026.
- Whether Hugging Face’s 1.7T safetensors count versus the paper’s 1.6T total / 49B active count is explained by the attached DSpark module; DeepSeek does not spell that out on the 0813 card.
- How long DeepSeek will keep the V4 Pro API after the 10 September 2026 continuation notice, which promised further notice of any change.

## Sources

1. [Change Log | DeepSeek API Docs](https://api-docs.deepseek.com/updates/) (primary)
2. [DeepSeek-V4-Pro GA Release | DeepSeek API Docs](https://api-docs.deepseek.com/news/news260813) (primary)
3. [deepseek-ai/DeepSeek-V4-Pro-0813 · Hugging Face](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro-0813) (primary)
4. [DeepSeek-V4 collection · Hugging Face](https://huggingface.co/collections/deepseek-ai/deepseek-v4) (primary)
5. [Models & Pricing | DeepSeek API Docs](https://api-docs.deepseek.com/quick_start/pricing/) (primary)
6. [Thinking Mode | DeepSeek API Docs](https://api-docs.deepseek.com/guides/thinking_mode/) (primary)
7. [DeepSeek V4 Pro 0813 - API Pricing & Benchmarks | OpenRouter](https://openrouter.ai/deepseek/deepseek-v4-pro-0813)
8. [DeepSeek V4 Pro 0813 (max) - Intelligence, Performance & Price Analysis | Artificial Analysis](https://artificialanalysis.ai/models/deepseek-v4-pro)
9. [DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence](https://arxiv.org/abs/2606.19348) (primary)

Author: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Multimodal](https://newruntime.com/topics/multimodal.md): Explore the multimodal topic hub.
- topic hub: [Openai](https://newruntime.com/topics/openai.md): Explore the openai topic hub.
- topic hub: [Tools](https://newruntime.com/topics/tools.md): Explore the tools topic hub.
- collection: [Analysis](https://newruntime.com/analysis.json): Open the latest authorial analysis.
- collection: [Source Ledger](https://newruntime.com/source-ledger.json): Inspect the public source evidence graph.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)