DeepSeek V4 Pro 0813 (Together AI)

DeepSeek-V4-Pro-0813 was rolled out quietly on 13 August 2026 across DeepSeek's app, web interface, and API, going generally available without a blog post or formal changelog, with the rollout emphasising "significantly enhanced agent capabilities." The article describes a new hybrid attention architecture together wit The post notes that DeepSeek's "significantly enhanced agent capabilities" phrasing was reportedly pulled from the company's site shortly after publication, a detail attributed to a South China Morning Post report. Independent benchmarkers reportedly found mixed results compared to DeepSeek's vendor-published numbers,

Tokenhot's technical guide covers DeepSeek-V4-Pro-0813 as the general-availability build released on August 13, 2026, with open weights published under the MIT license alongside cloud API endpoints. It characterizes the model as a fine-grained Mixture-of-Experts system with 1.6 trillion total parameters and 49 billion The article highlights a native 1-million-token context window supported by YaRN rotational position embeddings calibrated to maintain needle-in-a-haystack retrieval accuracy above 99.4 percent across the entire 1M token span, allowing full-codebase repositories and multi-hour telemetry logs to be processed without ext

DeepSeek released DeepSeek-V4-Pro-0813 on 13 August 2026 as the production, general-availability build of its flagship V4-Pro model, moving the existing deepseek-v4-pro endpoint off the April preview without requiring a code change from callers. The new build keeps a one million token context window and the same model On DeepSeek's own evaluation harness in minimal mode, V4-Pro-0813 scores 87.9 on Terminal-Bench 2.1, up from 72.1 for the April V4-Pro preview, with the article noting that DeepSeek's published numbers for rival models do not always match those labs' own figures. Alongside the model, DeepSeek open-sourced its Deepseek

China's National Supercomputing Internet announced on 15 August 2026 that it had made the official version of DeepSeek V4 Pro (DeepSeek-V4-Pro-0813) and the DeepSeek Harness agent framework available on its platform, providing users the ability to download the DeepSeek models and source code, conduct private deployment DeepSeek-V4-Pro-0813 is described as placing particular emphasis on enhanced agent capabilities, while DeepSeek Harness, open-sourced under the MIT license, adopts an "everything is a plugin" architecture that lets models, tools, skills, and sessions be freely replaced and recombined. Harness ships with four operating

On 12 August 2026, DeepSeek updated its official API pricing page to reflect the new DeepSeek-V4-Pro-0813 version behind the deepseek-v4-pro endpoint, ending a preview period that dated back to the model's 24 April 2026 debut, with OpenRouter listing the model the same day. There was no blog post, no changelog entry, a The article highlights that DeepSeek's published benchmark gains of 49.9 points on agent evaluations are vendor-reported on DeepSeek's own harness rather than independently verified, and that the same pricing page that confirmed GA carries a footnote warning of a "significant API price increase in the near future" with

DeepSeek-V4-Pro-0813 was released on Thursday, August 13, 2026, thirteen days after the V4-Flash-0731 sibling, as documented on AI Release Tracker. Benchmark coverage listed for the model includes BullshitBench v2, SWE-Bench Verified, Terminal-Bench 2.1, Toolathlon-Verified, BrowseComp, CyberGym, and two additional sui Rates shown on this tracker page were verified against api-docs.deepseek.com on August 18, 2026, reflecting post-August-16 pricing rather than the earlier preview-tier rates that other outlets cite. The page also lists a context-window specification and provider-availability fields describing how the same model can app

Unite.AI reports that the V4-Pro-0813 build appeared on August 12, 2026, as the general-availability release on OpenRouter's model page and that DeepSeek's own API documentation now lists DeepSeek-V4-Pro-0813 as the version behind the deepseek-v4-pro endpoint. The architecture is described as a mixture-of-experts syste The article carries over the preview-tier API pricing: $0.435 per million input tokens on a cache miss, $0.003625 per million on a cache hit, and $0.87 per million output tokens, with a one-million-token context window and a maximum output of 384,000 tokens, plus a concurrency limit of 500 for the Pro endpoint against

Chinese AI startup DeepSeek quietly released DeepSeek V4 Pro 0813 on August 13, 2026, a flagship reasoning model with 1.6 trillion total parameters, 49 billion active parameters, and a roughly 1M-token context window. The release quickly drew community attention for its reported cost-efficiency relative to leading closed models. DeepSeek reports benchmark jumps over the April preview, with Terminal Bench 2.1 rising from 72.1% to 87.9%, CyberGym from 52.7% to 83.3%, and DeepSWE from 12.8% to 62.7%. Listed API pricing is $0.435 per million input tokens and $0.87 per million output tokens, and open-source coding agent Cline highlighted the model as offering strong price-to-performance versus Claude Fable 5.