Ask HN: What default model do you use and why?

Hacker News Discussion ↗

AI power users are gravitating toward different “default” language models based on a trade-off between cost, speed, and capability rather than raw benchmark scores. Many favor cheaper, fast models like DeepSeek Flash, GLM 5.3, Gemini Flash, or Luna for day-to-day coding and automation, reserving more expensive frontier models such as Opus, Fable, Sol, or Astra for complex planning and review—or skipping them entirely due to verbosity and high token costs. A recurring theme is composing workflows from multiple models (and sometimes local open‑weight setups) to balance privacy, reliability, and budget, with some arguing that thoughtful harness and test design now matters more than chasing the “smartest” single model.

Model choices and use cases

  • Wide diversity; no single “default” winner.
  • Common “workhorses”: DeepSeek v4.1 Flash, GLM‑5.3‑Flash, GPT 5.6 Luna/Terra, Gemini 3.x Flash, Composer, Muse Spark.
  • Frontier models (Opus, Fable, Astra, Sol, Grok) mostly reserved for planning, hard bugs, or reviewing large codebases.
  • Some use cheap/fast models for everyday coding and logs, and switch up to Opus/Fable/Sol/Astra only when stuck.
  • A few avoid LLMs entirely.

Cost, token usage, and pricing frustration

  • Many optimize for price/performance; small personal budgets favor DeepSeek, GLM, Luna, Qwen, Gemma, local models.
  • Experiences with token limits differ wildly: some cannot exhaust large “20x”‑style plans, others burn through them in a day with multi‑agent factories and continuous devops.
  • Several feel high‑end Anthropic offerings are vastly overpriced for common dev tasks; others gladly pay for convenience and US‑based vendors.

Speed, capability, and verbosity

  • Fast response is highly valued; DeepSeek, GLM, Gemini Flash called “blazing fast” and “good enough” for most web/mobile or SWE work.
  • Significant backlash against newer Claude/Opus writing style: too verbose, flowery, cautious, and hard to parse; older Opus versions preferred.
  • GPT Sol/Luna, Astra, and some open‑weights praised for more concise, matter‑of‑fact output.
  • Some report frontier models pass tests yet still create hidden tech debt; harness/test design seen as critical.

Workflows and orchestration patterns

  • Common pattern: cheap model as main agent + smarter model as planner/reviewer/advisor.
  • Use of orchestrators, sub‑agents, and “software factories” is growing, but can explode token usage.
  • Several prefer “centaur” mode with tight human control rather than fully autonomous agents.

Local vs hosted, trust, and ethics

  • Some shift to local Qwen/Gemma/GLM/DeepSeek for privacy, credential handling, and to avoid perceived surveillance or policy issues.
  • Others are barred from Chinese models at work despite using them personally.
  • Debate over environmental impact and geopolitical alignment of US vs Chinese AI labs; no consensus.

Overall sentiment

  • Thread highlights extreme fragmentation: many viable defaults, strong personal preferences, and a sense that model choice often matters less than workflow, steering, and cost discipline.