DeepSeek V4 Pro 0813

In independent Artificial Analysis testing, DeepSeek V4 Pro 0813 scores 53 on the Intelligence Index compared to 57 for GLM 5.3 Flash, but generates at approximately 68 tokens per second versus 49 tokens per second for GLM 5.3 Flash. DeepSeek V4 Pro 0813 is described as stronger in terminal coding and repository-level The comparison places DeepSeek V4 Pro 0813 as a 1.6-trillion-parameter MoE with 49B active parameters and a one-million-token context window, contrasted against GLM 5.3 Flash's 320B total / 18B active parameters with multimodal support (text, images, video). DeepSeek V4 Pro 0813 wins on output speed, maximum output len

DeepSeek-V4-Pro-0813 was generally available on 13 August 2026 across the company's app, web interface, and API with a 1 million token context window and what the company described as significantly enhanced agent capabilities, language that was reportedly pulled from the company's own site shortly after. The release in The architecture walkthrough explains the hybrid attention mechanism layer by layer, paired with independent benchmark findings that put the model in context versus both the hype and the quiet walk-back of DeepSeek's own marketing claims. Independent benchmarking organizations found more mixed results than the launch l

DeepSeek released DeepSeek-V4-Pro-0813 to general availability on 13 August 2026, moving its flagship V4-Pro model out of preview while keeping the model name, parameter count, and 1 million token context window unchanged, so existing callers were migrated to the new checkpoint without code changes. On DeepSeek's own e Alongside the model, DeepSeek published its agent software, Deepseek Harness v0.1, as open source under the MIT license, and the 0813 build sits behind an Expert Mode toggle in the DeepSeek app and web interface. DeepSeek has not published weights for this build, leaving the April V4-Pro preview as the most recent vers

DeepSeek officially released its V4-Pro model on August 13, 2026, making it available across the company's app, web interface, and API after previewing it since April. The general availability version, designated DeepSeek-V4-Pro-0813, is positioned around agent capabilities — tasks where AI systems use tools, execute c The model handles a context window of up to 1 million tokens and produces outputs as long as 384,000 tokens, with the option to run in either thinking or non-thinking mode. DeepSeek also announced that the V4-Pro API has been updated to work with the OpenAI Responses API format out of the box and includes built-in supp

DeepSeek-V4-Pro-0813 is a text-only mixture-of-experts model with 1.6 trillion total parameters and 49 billion active parameters, supporting a one-million-token context window with up to 384K maximum output. It offers three thinking modes (non-thinking, Think High, and Think Max), native Responses API support, Anthropi Current API pricing per million tokens is $0.003625 for cache-hit input, $0.435 for cache-miss input, and $0.870 for output. Benchmark highlights include LiveCodeBench at 93.50, MMLU Pro at 87.50, and τ²-Bench Telecom at 96.20. The model is licensed under MIT for both code and weights, though the 0813 weights themselve

DeepSeek-V4-Pro-0813 was released on 13 August 2026, thirteen days after DeepSeek-V4-Flash-0731, and is covered across benchmarks including BullshitBench v2, SWE-Bench Verified, Terminal-Bench 2.1, Toolathlon-Verified, BrowseComp, and CyberGym. The model retains a one-million-token context window, and thinking and non- API pricing verified against api-docs.deepseek.com on 18 August 2026 lists off-peak rates at $0.66 input, $0.022 cached input, and $1.98 output per million tokens, with peak rates (01:00–04:00 and 06:00–10:00 UTC) at $1.32 input, $0.044 cached input, and $3.96 output per million tokens. The tracker confirms the model's

In a Codeforces competitive-programming rating snapshot dated 15 September 2026, DeepSeek V4 Pro 0813 scored 3206.0, placing second behind DeepSeek V4.1 Flash at 3471.0 and ahead of dots3-note Preview at 3056.0. The benchmark covers six models in the Coding category and confirms DeepSeek V4 Pro 0813 maintains a one-mil BenchLM.ai stores Codeforces as a display-only provider-table row because its rating scale is not a 0-100 percentage benchmark, and it is currently excluded from the scoring formula. The data was verified on September 15, 2026, with 33 confirmed releases in the prior 30 days. The broader top-10 range spans 655.0 score