DeepSeek V4 Pro
DeepSeek
DeepSeek is a fast-growing frontier model with a large and growing user base. We include it so the picture is not limited to a single provider's view of you.
Model details
- Model ID
deepseek-v4-pro- Knowledge cutoff
- 2025
Knowledge cutoff vs training data cutoff
Knowledge cutoff and training data cutoff are different dates: the knowledge cutoff is roughly how current a model's world knowledge is, while the training data cutoff is when its training data ends. Both shape what a model knows about you.
What we use this model for
We ask this model directly about your company to see how it describes you and which names it brings up in your space. It is one of several frontier models in our measurement set, so your visibility is read across the assistants people actually use rather than from a single model.
Recent updates for this model
-
today
DeepSeek released Harness v0.1, an open-source, plugin-first framework under an MIT license designed for building modular AI coding agents. The framework features a developer preview that allows users to extend or swap capabilities via plugins without modifying source code, while logging all model interactions.
- DeepSeek released Harness v0.1, an open-source, plugin-first framework for building AI coding agents under an MIT license.
- DeepSeek announced DeepSeek-V4-Pro, a flagship model specifically tuned for agentic workloads.
- DeepSeek Harness v0.1 features a plugin-based developer preview allowing users to swap or extend capabilities without changing source code.
- DeepSeek V4 Flash maintains a median cost per agent session of $0.04, compared to $0.10 for Gemini 3.7 Flash.
- DeepSeek launched Harness v0.1 with V4-Pro-0813, achieving high scores on Terminal Bench 2.1.
Sources: news@alphasignal.ai, superhumancode@news.codenewsletter.ai, welcome@openrouter.ai, astro@forwardfuture.ai, dan@tldrnewsletter.com, theneuron@newsletter.theneurondaily.com
-
1 days ago
DeepSeek released Harness v0.1, an open-source, plugin-first framework under an MIT license designed for building modular AI coding agents. The framework features a developer preview that allows users to extend or swap capabilities via plugins without modifying source code, while logging all model interactions.
- DeepSeek released Harness v0.1, an open-source, plugin-first framework for building AI coding agents under an MIT license.
- DeepSeek announced DeepSeek-V4-Pro, a flagship model specifically tuned for agentic workloads.
- DeepSeek Harness v0.1 features a plugin-based developer preview allowing users to swap or extend capabilities without changing source code.
- DeepSeek V4 Flash maintains a median cost per agent session of $0.04, compared to $0.10 for Gemini 3.7 Flash.
- DeepSeek launched Harness v0.1 with V4-Pro-0813, achieving high scores on Terminal Bench 2.1.
Sources: news@alphasignal.ai, superhumancode@news.codenewsletter.ai, welcome@openrouter.ai, astro@forwardfuture.ai, dan@tldrnewsletter.com, theneuron@newsletter.theneurondaily.com
-
2 days ago
DeepSeek released DeepSeek V4 Pro (version 0813) with a 1M token context window and 384K maximum output. The model is available via OpenRouter API and features pricing of $0.435 per 1M input tokens and $0.87 per 1M output tokens, including 50% off-peak discounts.
- DeepSeek released DeepSeek V4 Pro with a 1M token context window and 384K maximum output.
- DeepSeek V4-Pro-0813 is priced at $0.435 per 1M input tokens and $0.87 per 1M output tokens.
- DeepSeek V4 Pro and V4 Flash weights are available on Hugging Face, featuring 1.7T parameters.
- DeepSeek V4 Pro is available via OpenRouter API despite no official announcement page.
- DeepSeek V4 Pro offers 50% off-peak pricing.
- DeepSeek V4 Pro demonstrated notable gains on Terminal Bench.
Sources: simonwillison.net, news@alphasignal.ai, dan@tldrnewsletter.com
View the provider's model card Visit the provider
See how these models describe your company.
Run an analysis