GitHub vs Novelcrafter

Two threads run through this window. The first is model plumbing: default collections refreshed as Claude 4.5 Sonnet, Gemini 3 Pro, and GPT-5.1 arrive, a generic OpenAI-compatible provider so unsupported vendors work anyway, and an opt-in control to prefer or avoid reasoning across Scene Beats and Chats. The second is the writer's own data — Codex entries can be marked case-sensitive so a character named Will stops matching every sentence, and an entry can sit in several custom categories at once. Word export gained bullet and numbered lists. The feed stops in March 2026.

The model layer is treated as commodity and kept deliberately replaceable: users can clone collections, revert to previous ones, or point at any OpenAI-compatible endpoint, and the reasoning control is a preference rather than a hardcoded behavior. Novelcrafter's own investment goes into the Codex — the structured record of characters, places, and rules that gets fed to whichever model is in play — and into getting manuscripts out cleanly. That is a defensible split for a tool whose value is the writer's accumulated context rather than the generation itself.

Expect further Codex precision work, since case-sensitivity and multi-category membership are both fixes for the same failure — the wrong entries being pulled into context. What the entries do not show is whether the experimental reasoning support graduates out of the settings flag.