DeepSeek V4 Pro: the cheap AI that ties the best models on agent work.
It just went live. No launch video. No hype post. → Scored 87.9 on Terminal Bench (the top model got 88) → Reads 1 million tokens at once → Cache reads cost almost nothing → 90%+ cache hit rate on
$420 a month on AI. Then $20.
Same models Same quality Different path Claude Max, ChatGPT Pro, Cursor, Perplexity Most of that bill was convenience, not capability Free forever (no card): Google AI Studio · Gemini 2.5 Flash · ~1,500 req/day · 1M context Groq · Llama 3.3
I signed GPT, Claude launched a new model.
I signed Claude, Gemini and DeepSeek launched. That's it every week.
Just tested DeepSeek V4 Flash in the new DeepSeek Harness
I also discovered archify, a skill that turns a codebase into a visual architecture map. Gave it a try and asked it to explain the Harness itself. Pretty satisfied with the result for a first pass:
You have reached the end of the archive
All of deepseek