FULL STORY

DeepSeek V4 Pro: From Leaks to Launch and Backlash

Following leaked specs and pricing, DeepSeek V4 Pro officially launched with disruptive pricing but quickly faced community backlash over claims of being too similar to older versions.

2026-08-11 ~ 2026-08-13 · 4 episodes · 38 posts

Episode 1 · DeepSeek V4 Pro Surfaces in Rumors, Imminent Release Expected (2026-08-11, 5 posts)

Recent chatter across community and social platforms suggests the imminent release of DeepSeek V4 Pro. Developer @dejavucoder revealed that the official team is prepping and retaining the launch plan for this version. Shortly after, @jiayuanjy spotted a suspected deepseek-v4-pro-0813 API endpoint, adding credibility to the rumors. Multiple bloggers predict the model could drop as early as the evening of August 12 or 13.

已确认

  • 要点 接口曝光: Users discovered an API link for deepseek-v4-pro-0813, confirming the version number update.

尚未确认

  • 要点 发布时间: Online rumors suggest a release on August 12 (tomorrow) or the evening of August 13, but the exact time awaits official confirmation.
  • 要点 性能预期: Blogger @AlchainHust speculates that based on the recent performance of V4 Flash, the actual capabilities of the V4 Pro official release could approach the level of Opus 5 / Fable 5. However, this remains purely subjective speculation.

为什么重要

  • 要点 Following the massive success of DeepSeek V3, the market has exceptionally high expectations for the performance leap of its next flagship model, V4 Pro. If its performance truly approaches or matches top-tier competitor models, it will further reshape the competitive landscape of the current LLM industry.

Episode 2 · DeepSeek V4 Specs and Pricing Allegedly Leaked (2026-08-11, 5 posts)

Recent leaks from DeepSeek's API documentation and online rumors have revealed the core specs and suspected pricing for the upcoming V4 series, drawing massive community attention. While the capability parameters seem mostly clear, the pricing details are highly conflicting. All information awaits official confirmation.

Confirmed

  • According to leaked API docs shared by media and bloggers, DeepSeek is launching the V4 series, featuring deepseek-v4-flash and deepseek-v4-pro.
  • Regarding core specs, both models support a massive context length of up to 1M (one million) and a maximum output of 384K.
  • Netizen @赛博禅心 noted that the model name deepseek-v4-pro and its API node DeepSeek-V4-Pro might already be accessible, though no official announcement has been made.

Unconfirmed

  • There are two conflicting rumors regarding pricing. Leaks shared by @teortaxesTex indicate input at $2.40/1M tokens, output at $4.80/1M tokens, and cache hits at $0.20/1M tokens (a roughly 5.5x increase from current rates).
  • Another leak by @haider1 showing benchmark and pricing data suggests highly disruptive low prices: input at $0.435/1M tokens, output at $0.87, and cached input as low as $0.07 (refer to the original post for exact figures).

Why it matters

  • If the 1M context length holds true, DeepSeek V4 will join the industry's top tier for long-context processing. Whether the final pricing surges or remains highly competitive, it will directly impact the ongoing API price war in the LLM industry.

Episode 3 · DeepSeek V4 Pro Officially Released with Major Agent Upgrades and Disruptive Pricing (2026-08-12, 26 posts)

The official version of DeepSeek V4 Pro 0813 has been released and deployed across APIs and platforms like OpenRouter, featuring a 1 million token context window. Official and leaked benchmarks indicate significant improvements in Agent capabilities, though the community debates its actual performance. The model's highly disruptive pricing—offering single-task costs far below competitors—has triggered renewed industry focus on LLM inference costs.

已确认

  • DeepSeek V4 Pro 0813 official version is now available on DeepSeek API, the official chat platform, and third-party platforms like OpenCode Go and OpenRouter (@strangedell123, @zephyrz9, @aigclink, @op7418).
  • The model supports a 1 million token context, priced at $0.435/million tokens for input and $0.87/million tokens for output (@kimmonismus, @AccBalanced).
  • According to official reports, the model shows massive improvements over the preview version across multiple Agent benchmarks, such as DeepSWE reaching 62.7 (+49.9), Terminal Bench increasing to 87.9, and significant leaps in CyberGym scores (@scaling01, @ChrisGPT, @thesaraharminta).

尚未确认

  • Leaked test scores suggest explosive performance, even defeating current frontier models, but some data remains fully unverified by official sources (@op7418, @PMinervini).
  • There is community debate over the model's overall benchmark scores. @bindureddy revealed its scores lag behind Kimi K3 and Qwen; @kristoph also noted that AI labs often cherry-pick favorable benchmarks, making it difficult for users to judge true capabilities.
  • In hands-on comparisons, @vista8 stated that Grok 4.6 has a slightly better single-generation success rate than DeepSeek V4 Pro, but admitted the tests are subjective and random.

为什么重要

  • 极致性价比颠覆成本认知：Thanks to technologies like an ultra-small sparse KV Cache, its API pricing is exceptionally low. @ojasviyadav's tests indicate its single-task cost is 10 times cheaper than GLM5.2, and 72 times cheaper than Opus 4.8 and Fable 5. @XianbaoQIAN believes that if this model is open-sourced, it will have a disruptive impact.
  • 检验中国大模型后训练能力：@zephyrz9 points out that the model's performance will define the strength of Chinese AI labs in the post-training phase.
  • 加剧 API 价格战：@AccBalanced mentioned that since the release of the 0731 checkpoint, DeepSeek 4 Flash has already triggered an intense price war in the third-party market, and the community anticipates even fiercer competition driven by 4 Pro.

6 more related posts →

Episode 4 · DeepSeek V4-0813 Criticized for Showing No Improvement Over Flash Version (2026-08-13, 2 posts)

DeepSeek's newly released V4-0813 model faces backlash after evaluations show its performance is nearly identical to the previous V4-Flash-0731, with users criticizing the lack of meaningful improvements.