https://models.sutraworks.ai/lab/deepseek Changes and news for deepseek models: releases, price moves, deprecations. Fri, 21 Aug 2026 17:30:23 GMT https://eu.36kr.com/en/p/3948691750059394 98n7ya Fri, 21 Aug 2026 05:00:00 GMT For example, click the dialog box, select any folder on your local computer, which can be a newly created one or one containing specific project files; after selection, we can put forward requirements to DeepSeek Harness. Generally speaking, you can select the DeepSeek-V4-Pro model and set the reasoning level to Max. Although it costs a little more, it can also reduce the time we spend negotiating with AI. https://models.sutraworks.ai/m/deepseek/deepseek-v4-flash-vision-exp 0f092fad7fa1 Fri, 21 Aug 2026 00:00:00 GMT provider_added: DeepSeek V4 Flash Vision Exp via vercel https://models.sutraworks.ai/m/deepseek/deepseek-v4-flash 6eff15dbe370 Fri, 21 Aug 2026 00:00:00 GMT cost.input: 0.0826 → 0.08106; cost.output: 0.1652 → 0.16212; cost.cache_read: 0.01652 → 0.016212 https://models.sutraworks.ai/m/deepseek/deepseek-v4-flash-0731 7763dd2e80b3 Fri, 21 Aug 2026 00:00:00 GMT limit.output: 393216 → 384000 https://models.sutraworks.ai/m/deepseek/deepseek-v4-flash-vision-exp 794eeaab4750 Fri, 21 Aug 2026 00:00:00 GMT provider_added: DeepSeek V4 Flash Vision Exp via edenai https://models.sutraworks.ai/m/deepseek/deepseek-v4-pro-0423 85dbf20f6a46 Fri, 21 Aug 2026 00:00:00 GMT model_added: DeepSeek V4 Pro 0423 via ofox https://models.sutraworks.ai/m/deepseek/deepseek-v4-flash a06f9408bc8d Fri, 21 Aug 2026 00:00:00 GMT cost.input: 0.44 → 0.22; cost.output: 1.32 → 0.66; cost.cache_read: 0.014 → 0.007 https://models.sutraworks.ai/m/deepseek/deepseek-v4-flash-vision-exp c9627deeb41c Fri, 21 Aug 2026 00:00:00 GMT name: "DeepSeek: DeepSeek V4 Flash Vision Exp" → "DeepSeek V4 Flash Vision Exp"; family: "deepseek" → "deepseek-flash" https://models.sutraworks.ai/m/deepseek/deepseek-v4-pro d0e6d6958eee Fri, 21 Aug 2026 00:00:00 GMT cost.input: 1.32 → 0.66; cost.output: 3.96 → 1.98; cost.cache_read: 0.044 → 0.022 https://models.sutraworks.ai/m/deepseek/deepseek-v4-flash-0731 e0e585e9609a Fri, 21 Aug 2026 00:00:00 GMT limit.output: 393216 → 384000; cost.input: 0.14 → 0.08; cost.output: 0.28 → 0.18; cost.cache_read: 0.028 → 0.016 https://models.sutraworks.ai/m/deepseek/deepseek-v4-flash-vision-exp e6f21ba726f0 Fri, 21 Aug 2026 00:00:00 GMT family: "deepseek" → "deepseek-flash" https://models.sutraworks.ai/m/deepseek/deepseek-v4-flash-vision-exp f0163350a8af Fri, 21 Aug 2026 00:00:00 GMT model_added: DeepSeek V4 Flash Vision Exp via openrouter https://models.sutraworks.ai/m/deepseek/deepseek-v4-flash-vision-exp f29ea1aa3442 Fri, 21 Aug 2026 00:00:00 GMT provider_added: DeepSeek V4 Flash Vision Exp via ofox https://models.sutraworks.ai/m/deepseek/deepseek-v4-pro-0423 f696cbf56fc0 Fri, 21 Aug 2026 00:00:00 GMT provider_added: DeepSeek V4 Pro 0423 via merge-gateway https://tech-insider.org/ca/deepseek-v4-pro-vs-qwen3-8-max-vs-muse-glimmer-2026 ubiwlq Fri, 21 Aug 2026 00:00:00 GMT DeepSeek shipped the general-availability “0813” build of V4 Pro on August 13, 2026, replacing the April preview that had been running behind the same model ID. It is a mixture-of-experts model with roughly 1.6 trillion total parameters and about 49 billion active per token, built on a DeepSeekV4ForCausalLM architecture class, according to the model card DeepSeek published on Hugging Face. The headline engineering story is the attention mechanism: DeepSeek combined Compressed Sparse Attention [...] DeepSeek V4 Pro 0813 is available two ways: as hosted API access through DeepSeek’s own platform using OpenAI- or Anthropic-compatible request formats, or as a direct download from Hugging Face under the model ID `deepseek-ai/DeepSeek-V4-Pro-0813` for teams with the GPU capacity to self-host. Several router services, including OpenRouter, also proxy the hosted API for teams that want a single integration point across multiple model providers. [...] cache-hit price is close to free. Modality-wise, DeepSeek’s own spec pages describe V4 Pro 0813 as text-only, though at least one third-party comparison mentions experimental image reasoning in some deployments; treat vision support as unconfirmed until DeepSeek documents it directly. Weights are open under the MIT license and hosted on Hugging Face at `deepseek-ai/DeepSeek-V4-Pro-0813`, making it fully self-hostable for commercial use. https://tech-insider.org/gpt-5-6-vs-deepseek-v4-pro-0813-2026 1wtisdj Thu, 20 Aug 2026 12:10:05 GMT DeepSeek V4 Pro 0813, by contrast, is a single open-weight model, a refresh of the earlier DeepSeek V4 Pro release, dated to its August 13 build number. It is DeepSeek’s flagship reasoning and coding model, sitting above the cheaper V4 Flash tier the same way GPT-5.6 Sol sits above Luna. Where GPT-5.6 is API-only and fully closed, DeepSeek V4 Pro 0813 ships weights that developers can download, self-host, or run through third-party inference providers like OpenRouter, which changes the cost [...] DeepSeek’s strongest published reasoning gains, by contrast, cluster around legal and formal-proof tasks, plus general knowledge breadth. The Vals Index puts DeepSeek V4 Pro 0813’s overall score at 52.37%, a jump of 9.48 points over the prior DeepSeek V4 baseline’s 42.89%, moving it to #18 overall on that leaderboard. On the narrower BenchLM DeepSeek-only leaderboard, updated August 19, 2026, V4 Pro 0813 is ranked the top DeepSeek model with a score of 61.3, ahead of every earlier DeepSeek [...] Both models land in roughly the same context window class. GPT-5.6’s three tiers all carry approximately 1.05 million tokens of context, according to OpenAI’s model documentation pages. DeepSeek V4 Pro 0813’s context window is listed at 1.0 million tokens in the most recent cost calculator explicitly tied to the 0813 build, though it is worth noting that some older DeepSeek pricing guides still reference a smaller 256K window for a prior V4 Pro release, a discrepancy that likely reflects the https://tech-insider.org/how-to-set-up-deepseek-v4-pro-2026 3zgsoh Thu, 20 Aug 2026 10:09:36 GMT DeepSeek positions V4 Pro as its agentic and coding-first model, aimed squarely at complex reasoning chains, tool-calling workflows, and long-document analysis. The cheaper sibling, DeepSeek V4 Flash (`deepseek-v4-flash`, updated to the 0731 build on July 31, 2026), trades some reasoning depth for a 5x higher concurrency ceiling and a fraction of the per-token cost. We already covered Flash setup in a dedicated DeepSeek V4-Flash tutorial; this guide focuses on the Pro tier and the extra steps [...] DeepSeek V4 Pro’s positioning as an agentic model means tool calling (also called function calling) is where it’s meant to earn its higher price over Flash. The API follows the same tool-calling schema as OpenAI’s Chat Completions format, so if you’ve built agents against GPT models before, the pattern will look familiar. Here’s a minimal agent that gives the model access to a weather-lookup function and lets it decide when to call it: https://www.hpcwire.com/aiwire/2026/08/19/deepseek-open-sources-the-missing-layer-between-ai-models-and-agents 5okk9b Wed, 19 Aug 2026 16:47:46 GMT model. A sort of collection of software that determines whether an AI agent can actually get a job done. There are some hard numbers behind the release, too. Harness arrived alongside DeepSeek V4-Pro-0813, the latest version of the company’s flagship model, with much of the attention this time going to its agent capabilities. DeepSeek puts V4-Pro-0813 at 87.9 on Terminal Bench 2.1. It scored 74.1 on Toolathlon-Verified, 71.1 on DSBench-FullStack, and 67.2 on DSBench-Hard. These are DeepSeek’s [...] model. A sort of collection of software that determines whether an AI agent can actually get a job done. There are some hard numbers behind the release, too. Harness arrived alongside DeepSeek V4-Pro-0813, the latest version of the company’s flagship model, with much of the attention this time going to its agent capabilities. DeepSeek puts V4-Pro-0813 at 87.9 on Terminal Bench 2.1. It scored 74.1 on Toolathlon-Verified, 71.1 on DSBench-FullStack, and 67.2 on DSBench-Hard. These are DeepSeek’s [...] There are some hard numbers behind the release, too. Harness arrived alongside DeepSeek V4-Pro-0813, the latest version of the company’s flagship model, with much of the attention this time going to its agent capabilities. DeepSeek puts V4-Pro-0813 at 87.9 on Terminal Bench 2.1. It scored 74.1 on Toolathlon-Verified, 71.1 on DSBench-FullStack, and 67.2 on DSBench-Hard. https://gigazine.net/gsc_news/en/20260819-deepseek-v4-flash-agent z0gfa6 Wed, 19 Aug 2026 10:00:00 GMT While DeepSeek-V4 Flash achieved high scores in agent-based benchmarks, its success rate was lower on more challenging tasks. Taryn Prambu, a writer specializing in AI and cybersecurity, points out, 'This gap demonstrates why, in enterprise environments, the success or failure of a model depends not on the raw capabilities of the model itself, but on the 'orchestration' that brings multiple systems and tasks together. Even with the same model, the results can vary greatly depending on the [...] Technology analyst Karmi Levy stated that even after the price increase, DeepSeek remains 'far cheaper in every aspect' compared to competing models such as OpenAI, Anthropic, Google, and xAI, and emphasized the importance of appropriately allocating agent AI capabilities. For example, tasks like batch processing are often routine and repetitive, rather than requiring advanced technology, making it appropriate to use a cheaper and more efficient model. Levy said, 'Companies can consider [...] Sanchit Vil Gogia of Greyhound Research, a global technology research and consulting firm, said, 'What's important is whether the AI model's performance is sufficient to meet the actual workflows that companies operate, not whether it's the best at every benchmark. Companies need to determine which combination of model, harness, and provider will get the job done most securely and at the lowest cost.' https://www.forbes.com/sites/jonmarkman/2026/08/18/deepseek-releases-v4-pro-at-14x-the-price-of-its-cheapest-model 10d97hn Tue, 18 Aug 2026 00:00:00 GMT Illustration of DeepSeek AI on March 7, 2025. DeepSeek AI is a Chinese artificial intelligence company and the name of its conversational agent, which exploits a large language model. In early 2025, DeepSeek unveils a new model published in open source under the MIT license. Artificial intelligence. (Photo by Riccardo Milani / Hans Lucas / AFP via Getty Images) Hans Lucas/AFP via Getty Images [...] DeepSeek, once known for nearly free AI, has launched its V4 Pro model at prices up to 14 times higher than its budget offerings, charging $1.32/$3.96 per million tokens. This premium pricing, despite V4 Pro's marginal performance improvement over its cheaper V4 Flash, contradicts DeepSeek's prior narrative that AI was becoming too inexpensive. The company, which once caused a massive market dip for Nvidia by suggesting AI infrastructure was unnecessary, is now raising billions for data centers [...] DeepSeek spent the last eighteen months as the market's favorite proof that AI was getting too cheap to justify the money being spent on it. Last Friday the company released V4 Pro, its new flagship, at $1.32 per million input tokens and $3.96 per million output tokens. Its budget V4 Flash model costs $0.14 and $0.28. The company that made its name on nearly free intelligence is now charging up to fourteen times its own floor price for its best model.