DeepSeek V4 Pro 0813 (Cloudflare)

According to miraflow.ai's August 26, 2026 explainer, DeepSeek shipped DeepSeek-V4-Pro-0813 to general availability on August 13, 2026, in a notably quiet rollout that lacked a blog post or press release, with the lab emphasizing "significantly enhanced agent capabilities." The piece describes the underlying engineerin The same miraflow.ai post positions V4-Pro-0813 as a meaningful architecture shift rather than an incremental version bump, contrasting tight central attention with strided outer attention in its illustrative framing and explicitly citing the 1M-token context window and hybrid attention mechanism as the headline change

DeepSeek released DeepSeek-V4-Pro-0813 on 13 August 2026 as the general-availability build of its flagship model, moving the deepseek-v4-pro endpoint off the April preview without changing the model name, parameter count, or its one-million-token context window. The release ships alongside the MIT-licensed Deepseek Har On DeepSeek's own evaluation harness, DeepSeek-V4-Pro-0813 scores 87.9 on Terminal-Bench 2.1, up from 72.1 for the April V4-Pro preview, while the article notes DeepSeek's published numbers for rival models do not always match those labs' own figures. DeepSeek has not published weights for the 0813 build, so the April

TechTimes reported on August 13, 2026 that DeepSeek updated its official API pricing page on August 12, 2026 to reflect the new DeepSeek-V4-Pro-0813 version behind the deepseek-v4-pro endpoint, graduating from a preview that had been live since the model's April 24 debut, with OpenRouter listing the model the same day. The TechTimes piece also reconstructs the pricing history leading into GA: V4 Pro and the smaller V4 Flash launched as MIT-licensed open weights on April 24, 2026, an initial 75% discount became permanent on May 23 locking input at $0.435/M on cache miss and $0.003625/M on cache hit with output at $0.87/M, and V4 Flash

DeepSeek-V4-Pro-0813 is listed as an AI model released by DeepSeek on August 13, 2026, 13 days after the V4-Flash-0731 variant, with benchmark coverage spanning BullshitBench v2, SWE-Bench Verified, Terminal-Bench 2.1, Toolathlon-Verified, BrowseComp, and CyberGym among others. The page documents a 1M-token context win For developers integrating the model via Eden AI's Cloudflare-hosted route, the listing is useful as a concrete reference for DeepSeek's direct API pricing, context length, and benchmark menu for V4-Pro-0813, even though no Eden AI-specific markup or routing information is surfaced. The aggregator framing positions the

InferenceX's model card specifies DeepSeek-V4-Pro as a 1.6T total / 49B active parameter Mixture-of-Experts model with 61 layers and a native 1M-token context, arranged as 29 Heavily Compressed Attention layers alternating with 29 Compressed Sparse Attention layers, each routed through Top-6-of-385 expert selection wit The card positions V4-Pro as DeepSeek's flagship open-weights MoE, announced April 24, 2026 alongside the smaller V4-Flash, with 1M context as the default across DeepSeek services and dual Thinking/Non-Thinking modes. For developers integrating the Eden AI Cloudflare-hosted variant, this gives a precise reference for p