A recent Hacker News thread tracked GPT-5.6, Grok 4.5, Claude, and Muse Spark building the same four applications under identical prompts.
The exercise produced 130 points and 74 comments focused on measurable differences in code structure, error rates, and revision cycles.
Build-Off Task Breakdown
Each model received the same four specifications: a task manager with real-time sync, a minimal analytics dashboard, a file-upload API with validation, and a lightweight chat interface with persistence.
Prompts stayed fixed across runs. No model-specific tuning occurred.
Output Metrics from the Thread
Participants logged concrete results across 12 total model runs.
- GPT-5.6 completed all four apps with the fewest follow-up prompts (average 1.8 revisions).
- Grok 4.5 produced the longest initial code blocks but required 3.4 revisions on average.
- Claude delivered the cleanest TypeScript types in two of the four tasks.
- Muse Spark showed the fastest first-token response but the highest rate of incomplete functions.
How to Replicate the Test
Clone the four app specs from the original thread. Feed each model the exact prompt sequence while recording revision count and final test coverage.
Run the same unit-test suite on every output. Track tokens used and wall-clock time per iteration.
Observed Tradeoffs
GPT-5.6 produced the most consistent folder structures and fewer runtime errors. Grok 4.5 generated richer feature sets on the first pass but introduced more merge conflicts during integration.
Claude excelled at strict type safety yet sometimes over-engineered simple endpoints. Muse Spark stayed fastest for prototypes but left more TODO comments.
Who Should Run Similar Tests
Teams selecting a primary coding model benefit from repeating the four-app exercise on their own stack. Solo developers already satisfied with one provider can skip the overhead.
Organizations evaluating cost per successful deployment should weight revision count more heavily than raw generation speed.
Model Selection Verdict
The thread data indicates GPT-5.6 currently leads on end-to-end reliability for small-to-medium internal tools, while Grok 4.5 remains competitive when maximum feature density on the first attempt matters more than polish.
Bottom line: Re-running the exact four-app prompts on current frontier models gives developers the clearest signal for production use.
Early comments note that prompt length and example count affected outcomes more than model size alone.
Top comments (0)