Aikido's cyber-capability benchmark burned 11.7 billion tokens across 10 AI models, each running 32 fresh off-the-shelf vulnerabilities three times. DeepSeek V4 Pro 0813 was the top performer, pooling three runs to find 28 of 32 vulnerabilities, with three DeepSeek Pro runs costing approximately $295 and outperforming The benchmark reports that single runs miss much of the breadth, but pooling across runs fills that gap: DeepSeek Pro finds 17 vulnerabilities on its first pass but 28 across three. It concludes that open-source models now outperform the public frontier on pooled vulnerability recall, with DeepSeek V4 Pro topping every

DeepSeek officially launched V4-Pro on August 13, 2026 across its app, web interface, and API after months in preview since April, according to QZ. The GA version, designated DeepSeek-V4-Pro-0813, focuses on agent capabilities, scoring 87.9 on Terminal Bench 2.1, 62.7 on DeepSWE, and 61.5 on NL2Repo among other agent-f The QZ report also documents that the V4-Pro API has been updated for OpenAI Responses API compatibility with built-in Codex integration, and that thinking-effort levels for both V4-Pro and V4-Flash now expand to three settings (low, high, max). A price increase taking effect August 16 introduces peak and off-peak bill

DeepSeek released DeepSeek-V4-Pro-0813 on August 13, 2026 as the GA version of its flagship model, thirteen days after DeepSeek-V4-Flash-0731, according to AI Release Tracker's model page. The page documents benchmark results spanning BullshitBench v2, SWE-Bench Verified, Terminal-Bench 2.1, Toolathlon-Verified, Browse The DeepSeek-V4-Pro-0813 page is sourced from DeepSeek's own api-docs.deepseek.com rather than any SenseNova serving endpoint, attributing the model to DeepSeek as the provider and answering API requests. It does not establish any SenseNova-specific hosting, routing, or availability claim, so it is useful strictly as t

DeepSeek's official API changelog documents the GA release of DeepSeek-V4-Pro on August 13, 2026 across the APP, Web, and API, with the API model name remaining deepseek-v4-pro. The update lists substantially enhanced agent-capability results: HLE (wo/w tools) 42.7/60.0, Terminal Bench 2.1 87.9, NL2Repo 61.5, Cybergym The same changelog records that the DeepSeek API now natively supports the OpenAI Responses API format with specific Codex adaptation via a one-click configuration script, and that V4-Pro and V4-Flash thinking modes now expose three effort levels (low, high, max). It also covers the separate August 21 release of DeepSe