What This Is
This week, a developer tested Claude Opus 5 and GPT 5.6 Sol on the same problem — and the results are telling: GPT, which reached for the web, got it wrong first (blaming IPv6), while Opus, reasoning on its own, nailed it immediately. The issue: OpenCode CLI launched locally and hung indefinitely on "loading." Opus locked onto the "launch directory" as the key variable right away; GPT only arrived at the truth after three prompts from the developer — OpenCode runs a git snapshot scan from the user's home directory, which jams because the directory is too large (210,000+ files, 12 junctions).
Industry View
What caught our editorial eye was the developer's takeaway: every model that searched the web failed; every one that relied on its own reasoning succeeded. On its first attempt, GPT pulled "IPv6 issues other users hit in the official issue tracker" — similar symptoms, fundamentally different root cause. This is the textbook trap of retrieval-augmented generation (RAG, i.e., letting AI look things up on the fly before answering). Another layer of risk: AI coding assistants are still less sensitive to "environment differences" than human engineers. Attributing problems to the network layer or the version layer is easier than blaming "directory structure," so these models prefer to "borrow strength" rather than "grind it out."
Impact on Regular People
- Enterprise IT deployment: AI coding assistants still make unstable root-cause judgments in unfamiliar environments; blind trust can mask real failure points.
- Workplace troubleshooting: prompts should explicitly state "reason first, then retrieve," otherwise AI gets derailed by look-alike cases the moment it touches the network.
- Consumer market selection: "can it think for itself" is more worth paying for than "can it look things up."