Chinese AI lab DeepSeek’s last model release was V3.2 (and V3.2 Speciale) last December. They just dropped the first of their hotly anticipated V4 series in the shape of two …
API rate is absurdly cheap at least until the end of the month, 75℅ off list currently
Anecdotally, on recent code instead of pelicans, it’s noticeably worse than Western SOTA but strong prompts and a more competent operator than me would help close the gap
IMO the valuation shock when the first deepseek models came out were an under-reaction: Anthropic and OpenAI are months ahead, not years
I know at least two people who’ve been using Claude for coding and are cancelling their subscriptions after driving DeepSeek v4 for a week. My experience is that they’re fairly comparable. My workflow has been to get it to make phased plans in small verifiable steps, tell it to use test driven development where it writes the spec first and then implements against it, and to do code reviews at the end. I find that tends to work pretty reliably. Another very helpful trick is to ask it to make mermaidjs diagrams you can look at and review for the initial design. Then you can nail down what it’s actually going to be doing that way before it starts implementing anything.
API rate is absurdly cheap at least until the end of the month, 75℅ off list currently
Anecdotally, on recent code instead of pelicans, it’s noticeably worse than Western SOTA but strong prompts and a more competent operator than me would help close the gap
IMO the valuation shock when the first deepseek models came out were an under-reaction: Anthropic and OpenAI are months ahead, not years
I know at least two people who’ve been using Claude for coding and are cancelling their subscriptions after driving DeepSeek v4 for a week. My experience is that they’re fairly comparable. My workflow has been to get it to make phased plans in small verifiable steps, tell it to use test driven development where it writes the spec first and then implements against it, and to do code reviews at the end. I find that tends to work pretty reliably. Another very helpful trick is to ask it to make mermaidjs diagrams you can look at and review for the initial design. Then you can nail down what it’s actually going to be doing that way before it starts implementing anything.