Thread

1. monksy · 1790970968 · focus · HN ↗

   GLM5.3-flash has been fantastic for me to make minor fixes in ambigious ways. "Fix x feature, whats going wrong. " It does the job.

   1. surgical_fire · 1790971245 · focus · HN ↗

      GLM-5.3-flash is my implementation model after GLM-5.3 writes the plan.

      It's an excellent workhorse. When I am running out of my GLM quota I switch GLM-5.3-flash to DS-4.1-flash.

      1. finnjohnsen2 · 1790978270 · focus · HN ↗

         Do you switch model mid session, or do you use subagent to do the switch after planning?

         1. surgical_fire · 1790984123 · focus · HN ↗

            Neither, I use different sessions for each step.

            1. finnjohnsen2 · 1791036656 · focus · HN ↗

               So your planner leaves markdown files I assume...?

               1. surgical_fire · 1791039723 · focus · HN ↗

                  Correct. I customized my harness so agents communicate with each other through markdown files.

                  I find that it makes coding and review at the same time less prone to errors and cheaper

         2. drob518 · 1791035845 · focus · HN ↗

            I sometimes switch mid-session. Depends what I’m doing. Sometimes a model is slow or it’s not giving me what I want and I don’t want to dump the context. But if there’s a natural break, I’ll start a new session with a new model.
      2. monksy · 1790988654 · focus · HN ↗

         Same. It's barely even touching the credit I have on openrouter, it's great.
      3. drob518 · 1791035711 · focus · HN ↗

         Those are my two workhorses as well.
   2. girvo · 1790992346 · focus · HN ↗

      Between GLM 5.3 Flash on my legacy Z.ai coding plan, and Qwen 3.8 Flash locally on my DGX Spark-like, I'm barely using my Anthropic/OpenAI subscriptions, likely to cancel them soon.

      1. vardalab · 1790996013 · focus · HN ↗

         It's been slow like molasses on the coding plan. I ended up using it more on fireworks. But DeepSeek is so much cheaper because the cache cost is much better.

         1. girvo · 1790997014 · focus · HN ↗

            I’m kind of lucky that most of the time I don’t have to use it through peak times so it’s not that bad speed wise

            The legacy plan I have is so good as to be basically unlimited usage for my workloads, so I’m kind of stuck with it til they stop renewing it haha

         2. drob518 · 1791035658 · focus · HN ↗

            Cache cost is the primary metric, imo. Most people look at output cost, but you only pay that once. Cache cost you pay every turn and it grows each turn as the context length grows.
         3. pmoriarty · 1791044550 · focus · HN ↗

            The biggest problem with DeepSeek 4.1 Flash is all that it hallucinates a lot.

            [1] - <a href="https:&#x2F;&#x2F;www.youtube.com&#x2F;watch?v=P4dTq4X8bqk" rel="nofollow">https:&#x2F;&#x2F;www.youtube.com&#x2F;watch?v=P4dTq4X8bqk

Unofficial Hacker News client; not affiliated with Y Combinator.