DeepSeek V4 Pro

DeepSeek's official news page, dated September 9, 2026, introduces DeepSeek-V4.1-Flash as the smallest model in its new architecture family with native visual understanding, describing an asymmetric 552B-parameter MoE design with a new Causal Encoder–Decoder architecture using just 8B active parameters for input and 16 The same announcement states that DeepSeek is phasing out V4-Pro: starting at 04:00 UTC on September 14, 2026, all deepseek-v4-pro requests will route to V4.1-Flash at V4.1-Flash rates, a state that will continue until V4.1-Pro launches. The V4-Flash and V4-Flash-Vision-Exp models are listed as retired, with their lega

Miraflow's technical write-up describes DeepSeek-V4-Pro-0813, released August 13, 2026, as a general-availability build of DeepSeek's flagship model featuring a new hybrid attention architecture, an unusually deep stabilized residual stack, and a 1 million-token context window, and frames the rollout as a substantial e Miraflow walks through the V4-Pro-0813 architecture mechanism by mechanism and benchmarks it against independent third-party tests, concluding that the hybrid attention design and stability techniques deliver measurable gains on agentic and reasoning evaluations even where DeepSeek's headline numbers do not always lead

Aikido's independent benchmark, published August 21, 2026, ran 10 models through 32 fresh off-the-shelf vulnerabilities three times each (96 runs total, 11.7 billion tokens) to compare cybersecurity vulnerability-discovery capability. DeepSeek V4 Pro 0813 was the top performer on pooled vulnerability recall, finding 28 This is one of the few independent third-party evaluations of the V4 Pro 0813 GA build, providing original empirical evidence that the model competes with top closed frontiers on agentic security tasks and quantifying the cost of doing so. It complements DeepSeek's vendor-reported benchmarks with external validation on

DeepSeek V4 Pro 0813 is a 1.6-trillion-parameter Mixture-of-Experts model with 49 billion active parameters per token, a 1-million-token context window, 384K maximum output, and MIT-licensed open weights, according to a detailed independent review published August 18, 2026. The 0813 build is framed primarily as a servi DeepSeek reports 87.9 on Terminal-Bench 2.1, but CoderSera measured 54.68% on a neutral harness under the same conditions, while V4 Flash reached 67.04% — a gap that raises reproducibility questions. Artificial Analysis assigns the 0813 release an Intelligence Index score of 53 and an AA-Omniscience score of 0.83, plac

DeepSeek officially launched its V4-Pro model on August 13, 2026, making the general-availability version, designated DeepSeek-V4-Pro-0813, available across DeepSeek's app, web interface, and API after several months in preview. According to the report, V4-Pro-0813 emphasizes agent capabilities, with DeepSeek-published The launch introduced API-level updates including native compatibility with the OpenAI Responses API format, built-in Codex integration support, and three thinking-effort settings (low, high, max) for V4-Pro and V4-Flash. Alongside the GA release, DeepSeek announced new pricing taking effect August 16, 2026, with V4-Pr

BenchLM's page for DeepSeek V4 Pro 0813 aggregates published evidence for the model and reports a 1 million-token context window alongside an 80 tok/s throughput figure and a 26.87-second first-token latency. Its category table lists Agentic at 55.0 (rank 37 of 154, 76th percentile, 11/11 verified benchmarks), Coding a The aggregator further characterizes its evidence base by separating verified from provisional rows and by reporting 40 published rows across the tracked catalog of 497 models, with the model's blended, input, output, and cached pricing fields listed alongside but treated here as serving-side context rather than model

DeepSeek-V4-Pro-0813 was released on August 13, 2026, succeeding the April preview of V4 Pro. The AI Release Tracker page documents the exact -0813 variant with off-peak pricing of $0.66 input / $0.022 cached / $1.98 output per million tokens and peak pricing (01:00–04:00 and 06:00–10:00 UTC) of $1.32 / $0.044 / $3.96, According to the tracker, the page is attributed to DeepSeek as the model publisher, with provider availability shown. The release is positioned 13 days after DeepSeek-V4-Flash-0731, confirming the model sequence within the V4 family. As a third-party tracker, the figures ultimately derive from DeepSeek's own release m