$大肥鱼 DeepSeek R1 After 3 Months: The Best Value in AI Right Now
I've used DeepSeek R1 for months on real work — proposals, spreadsheets, emails, technical translation. Here's my honest take: it won't replace ChatGPT for everything, but for logic-heavy work it's the best value in AI right now, and more people should be paying attention.
DeepSeek R1 is a reasoning model — it works through a chain of thought before answering, so it shines on multi-step logic. It's excellent at debugging Excel/Google Sheets formulas, pinpointing errors in nested VLOOKUPs and explaining why. It traces contradictions across contract clauses that skimming by eye would miss. It plans well under multiple constraints without over-engineering. It handles light technical work like Apps Script, regex, and payroll formulas accurately, and translates technical documents with solid context. On benchmarks, the May R1-0528 update pushed its AIME 2025 math score from 70% to 87.5%, and on Codeforces it outperformed 96.3% of human participants.
It has weak spots: marketing copy comes out correct but dry, it's slow when you need instant replies, real-time data can be outdated unless verified, and in sensitive contexts like medical, reports note it can produce inconsistent conclusions — so verify anything critical.
But the cost argument is hard to ignore. Training DeepSeek-R1-Zero is estimated at just around $294,000, the web chat is essentially free, and the API is about 96% cheaper than OpenAI o3.
Treat DeepSeek as a low-cost logic specialist: give it spreadsheets, contracts, code, and technical translation. Keep another model — or yourself — for tone, speed, or absolute accuracy. If you're paying for AI mainly to process logic and data, DeepSeek R1-0528 deserves a serious look.
I've used DeepSeek R1 for months on real work — proposals, spreadsheets, emails, technical translation. Here's my honest take: it won't replace ChatGPT for everything, but for logic-heavy work it's the best value in AI right now, and more people should be paying attention.
DeepSeek R1 is a reasoning model — it works through a chain of thought before answering, so it shines on multi-step logic. It's excellent at debugging Excel/Google Sheets formulas, pinpointing errors in nested VLOOKUPs and explaining why. It traces contradictions across contract clauses that skimming by eye would miss. It plans well under multiple constraints without over-engineering. It handles light technical work like Apps Script, regex, and payroll formulas accurately, and translates technical documents with solid context. On benchmarks, the May R1-0528 update pushed its AIME 2025 math score from 70% to 87.5%, and on Codeforces it outperformed 96.3% of human participants.
It has weak spots: marketing copy comes out correct but dry, it's slow when you need instant replies, real-time data can be outdated unless verified, and in sensitive contexts like medical, reports note it can produce inconsistent conclusions — so verify anything critical.
But the cost argument is hard to ignore. Training DeepSeek-R1-Zero is estimated at just around $294,000, the web chat is essentially free, and the API is about 96% cheaper than OpenAI o3.
Treat DeepSeek as a low-cost logic specialist: give it spreadsheets, contracts, code, and technical translation. Keep another model — or yourself — for tone, speed, or absolute accuracy. If you're paying for AI mainly to process logic and data, DeepSeek R1-0528 deserves a serious look.