DeepSeek
DeepSeek V4 Pro
The Pro tier for hard reasoning, not everyday coding
What it is
DeepSeek’s Pro tier: a 1M-token context window, thinking on by default and open weights. No vision. Peak pricing is $0.66/$1.98 per 1M tokens; off-peak is half.
NLT take
Reach for Pro when Flash gets the reasoning wrong, not by default. The price is roughly three times Flash for the same context, and the weights are a real self-host option.
Before you switch
No image input, unlike Flash. If your agent reads screenshots, keep that step on another model.
Cost for a real task
The same task on every model: 50,000 input tokens and 10,000 output tokens, at the base tier.
Task cost
$0.053
50K input + 10K output tokens
- Input, per 1M
- $0.66
- Output, per 1M
- $1.98
- Cached input, per 1M
- $0.044
Peak rates. Off-peak is half: $0.33 input / $0.99 output.
Pricing source: DeepSeek models and pricing
Specs
- Maker
- DeepSeek
- Licence
- Open weights
- Released
- August 13, 2026
- Context window
- 1,000,000 tokens
- Max output
- 384,000 tokens
- Input / output
- Text → Text
Also known as: deepseek-v4-pro · DeepSeek-V4-Pro-0813
Watch it in action
Hands-on videos from creators we trust. They open on YouTube; we do not host or re-upload them.
DeepSeek V4 Pro Is HERE – Is THIS the BEST Open Model Yet?Bijan BowenWhy watch: A broad test that separates the Pro tier from Flash.Watch on YouTube(opens YouTube in a new tab)
Deepseek strikes againAI SearchWhy watch: A shorter, sceptical look at what the Pro tier changes.Watch on YouTube(opens YouTube in a new tab)
From our coverage
Articles about this model or its family. A “Background” tag means we have not covered this exact release yet.
- BackgroundMay 2, 2026DeepSeek V4: Open-Weight Frontier at 1/7 the CostDeepSeek V4 ships 1.6T MoE open weights with 1M-token context: 80.6% on SWE-bench Verified at $1.74/$3.48 per million — roughly 1/7 the output cost of Claude Opus 4.7.
- BackgroundAugust 19, 2026DeepSeek Harness: Free Agent Loop, Costlier Tokens 2026DeepSeek open-sourced its agent harness under MIT and it passed 161,000 GitHub stars in six days, then raised API prices right where agent loops hurt most.
Sources & model links
Every number on this card traces back to one of these links. The maker page proves the positioning, the pricing page proves the rates, the Hugging Face repo proves the weights exist, and the Ollama page proves a local tag.
Official
- DeepSeekVerifies: maker
- Models and pricingVerifies: price · specs
Open weights
- Hugging Face weightsVerifies: weights · licence
Run locally
- Run with Ollama (tag cloud)Verifies: local runThe published tag is cloud.