Dennis on Nostr: Still works well for me: Plan and build initial versions/milestones with potent ...

Still works well for me: Plan and build initial versions/milestones with potent models (e.g. GLM), then switch to cost-efficient ones for fine-grained tasks and executing. In that category DeepSeek v4 Flash works suprisingly good for it being so cheap.

quoting
nevent1q…d9qz

Closing in on my monthly OpenCode Go limits I started using cheaper models to see what they are good for: As longs as used with a narrow scope (e.g. per feature or improving test coverage) cost-efficient models like DeepSeek v4 Flash get the job done, though they might require one or two additional iteration steps.

At least for the stuff I'm building I'd say one doesn't have to spent lots of money for inference to get good results. Also I noticed how prompting efficiently forced me to become a better product steward, thinking through features in terms of goals rather than implementation details.