--- title: "xiun xiao (@xiunxiao) | Hashnode" description: "xiun xiao (@xiunxiao) writes on Hashnode. Read their latest articles and follow their work." url: "https://hashnode.com/@xiunxiao" --- @xiunxiao # xiun xiao @xiunxiao · Joined November 2026 ### [Image generation API compared: 11 models, the same 3 prompts, real outputs](https://xiunxiao.hashnode.dev/image-generation-api-compared-11-models-the-same-3-prompts-real-outputs?utm_source=hashnode&utm_medium=feed) GPT Image 2.5, Seedream, Qwen Image, FLUX, SDXL Turbo and more on the same three prompts: text rendering, composition, latency and price per image. 9 min read · [xiunxiao.hashnode.dev](https://xiunxiao.hashnode.dev/image-generation-api-compared-11-models-the-same-3-prompts-real-outputs?utm_source=hashnode&utm_medium=feed) [8h ago](https://hashnode.com/posts/image-generation-api-compared-11-models-the-same-3-prompts-real-outputs/6af2731c71d39c857cdbd713) ### [Kimi K3.1 Pro Highspeed vs Pro: we measured the real speedup](https://xiunxiao.hashnode.dev/kimi-k3-1-pro-highspeed-vs-pro-we-measured-the-real-speedup?utm_source=hashnode&utm_medium=feed) Same model, twice the price. Highspeed generated text 2 to 3x faster but did not start sooner; a coding agent finished 2.5x faster for 1.5x the cost. 9 min read · [xiunxiao.hashnode.dev](https://xiunxiao.hashnode.dev/kimi-k3-1-pro-highspeed-vs-pro-we-measured-the-real-speedup?utm_source=hashnode&utm_medium=feed) [8h ago](https://hashnode.com/posts/kimi-k3-1-pro-highspeed-vs-pro-we-measured-the-real-speedup/6af273178110798ac6c1d6a6) ### [DeepSeek V4.1 Flash vs Claude Haiku 5.5: a 17-task benchmark of the two cheapest tiers](https://xiunxiao.hashnode.dev/deepseek-v4-1-flash-vs-claude-haiku-5-5-a-17-task-benchmark-of-the-two-cheapest-tiers?utm_source=hashnode&utm_medium=feed) The two cheapest tiers on the same 17 coding, extraction and reasoning tasks, three runs each: pass rate, latency and cost per 1,000 tasks. 6 min read · [xiunxiao.hashnode.dev](https://xiunxiao.hashnode.dev/deepseek-v4-1-flash-vs-claude-haiku-5-5-a-17-task-benchmark-of-the-two-cheapest-tiers?utm_source=hashnode&utm_medium=feed) [8h ago](https://hashnode.com/posts/deepseek-v4-1-flash-vs-claude-haiku-5-5-a-17-task-benchmark-of-the-two-cheapest-tiers/6af273133f36efab05c00df0) ### [Claude Opus 5.5 vs GPT-6.1 Sol for coding agents](https://xiunxiao.hashnode.dev/claude-opus-5-5-vs-gpt-6-1-sol-for-coding-agents?utm_source=hashnode&utm_medium=feed) Same coding agent, six tasks with hidden tests, two runs per model. Both passed every test; Opus took about half the time at about the same cost. 10 min read · [xiunxiao.hashnode.dev](https://xiunxiao.hashnode.dev/claude-opus-5-5-vs-gpt-6-1-sol-for-coding-agents?utm_source=hashnode&utm_medium=feed) [8h ago](https://hashnode.com/posts/claude-opus-5-5-vs-gpt-6-1-sol-for-coding-agents/6af2730fd0e3db4dc4c07bbc) ### [Gemini 4 Argon vs Claude Opus 5.5 for coding agents](https://xiunxiao.hashnode.dev/gemini-4-argon-vs-claude-opus-5-5-for-coding-agents?utm_source=hashnode&utm_medium=feed) Same coding agent, six tasks with hidden tests, two runs per model. Both passed every test; Gemini took 1.6x as long and cost 1.5x as much per task. 10 min read · [xiunxiao.hashnode.dev](https://xiunxiao.hashnode.dev/gemini-4-argon-vs-claude-opus-5-5-for-coding-agents?utm_source=hashnode&utm_medium=feed) [8h ago](https://hashnode.com/posts/gemini-4-argon-vs-claude-opus-5-5-for-coding-agents/6af2730a806cce4909d7e78d)