DeepClaude: Anthropic's UX, DeepSeek's prices
DeepClaude jumped onto HN's front page overnight and pulled 467 upvotes in 11 hours. 547 stars in two days. The pitch is one sentence that explains why people care: Claude Code's autonomous coding loop, but the API calls go to DeepSeek V4 Pro instead of Anthropic.
The math is brutal. Anthropic Max sells at $200/month with usage caps. DeepSeek V4 Pro charges $0.87 per million output tokens, with automatic context caching that drops to $0.004/M after the first request. The author runs the numbers on his README: light user (10 days/mo) drops from $200 to $20. Heavy user (25 days/mo) drops from $200 to $50. So 75 to 90 percent off whatever Anthropic is charging you.
What stays the same is the body. File editing, bash execution, subagent spawning, multi-step autonomous loops. Anthropic's harness is the part everyone agrees is the best in the business. DeepSeek V4 Pro is a frontier coding model that scores 96.4% on LiveCodeBench. So you keep Anthropic's product experience but pay frontier-Chinese-model prices.
This is the second DeepSeek-stack data point in 24 hours, after DeepSeek-TUI hit GitHub trending yesterday. Different angle though. DeepSeek-TUI is a standalone Rust TUI that ignores Claude Code entirely. DeepClaude wraps Claude Code and routes its API calls. Two paths to the same conclusion. The commercial coding agent stack does not have to be Anthropic's. The harness layer was always horizontal, and the model layer is now competitive. Both layers swappable.
The repo is here: https://github.com/aattaran/deepclaude. MIT licensed. The core trick is that DeepClaude only sets the Claude-Code config per session, then restores it on exit, so it's nondestructive to your existing setup. Worth ten minutes of testing.
← Back to all articles
The math is brutal. Anthropic Max sells at $200/month with usage caps. DeepSeek V4 Pro charges $0.87 per million output tokens, with automatic context caching that drops to $0.004/M after the first request. The author runs the numbers on his README: light user (10 days/mo) drops from $200 to $20. Heavy user (25 days/mo) drops from $200 to $50. So 75 to 90 percent off whatever Anthropic is charging you.
What stays the same is the body. File editing, bash execution, subagent spawning, multi-step autonomous loops. Anthropic's harness is the part everyone agrees is the best in the business. DeepSeek V4 Pro is a frontier coding model that scores 96.4% on LiveCodeBench. So you keep Anthropic's product experience but pay frontier-Chinese-model prices.
This is the second DeepSeek-stack data point in 24 hours, after DeepSeek-TUI hit GitHub trending yesterday. Different angle though. DeepSeek-TUI is a standalone Rust TUI that ignores Claude Code entirely. DeepClaude wraps Claude Code and routes its API calls. Two paths to the same conclusion. The commercial coding agent stack does not have to be Anthropic's. The harness layer was always horizontal, and the model layer is now competitive. Both layers swappable.
The repo is here: https://github.com/aattaran/deepclaude. MIT licensed. The core trick is that DeepClaude only sets the Claude-Code config per session, then restores it on exit, so it's nondestructive to your existing setup. Worth ten minutes of testing.
Comments