2026-07-06 OpenRouter Interceptor Curated Model List
What I set out to do
Trim the per-run model dropdown on the OpenRouter Interceptor’s reload/analyze query cards. The recent reload feature (see 2026-07-04 OpenRouter Interceptor Queries UX Features) mapped OpenRouter’s entire /models catalog (340 entries, price-sorted) into the <select>, which was unusable. Goal: cut it to a reasonable curated set.
What I actually did
Traced the source: the dropdown flows through toModelOptions in src/utils/openrouter-models.ts, the single funnel feeding the query select, the custom-analysis modal, and the options page. Added a CURATED_MODEL_IDS allowlist and filtered there (DRY across all three consumers), with a safety net: if nothing matches (stale allowlist vs. a changed catalog) it returns the full list rather than an empty dropdown. Wrote the tests first (TDD) — 12 passing.
The interesting part was choosing the models. First pass I picked “best analysis” flagships (GPT-5.4, Claude Opus 4.8, Gemini Pro). That was wrong for the use case: the feature re-runs captured roleplay traffic, so a filtered flagship refuses or sanitizes NSFW/dark content and is useless on the actual captures. Reframed around cost/performance for roleplay with permissive (low content-filter) models.
Grounded the choice in real data per my LLM Comparison Sources methodology — the OpenRouter roleplay collection (revealed preference, Karpathy’s un-gameable signal) plus partial UGI willingness scores and live pricing. That surfaced a DeepSeek-dominated, mostly cheap/open-weight tier. Final call after the user asked for cheaper options: dropped the entire Claude/GPT-5/Gemini-Pro flagship tier, landed on 10 permissive models all under ~1.20 out, headed by deepseek/deepseek-v4-flash (0.18, the #1 RP model) and a free Nemotron-3 Ultra.
What was striking
“Best model” is entirely use-case-relative. The analysis-optimal list and the roleplay-optimal list barely overlap — content-filter posture flips the ranking, and the cheapest models (DeepSeek, open-weight) are the best RP picks, not a compromise. Revealed preference (OpenRouter RP token usage) cut through my stale training-cutoff knowledge of mid-2026 models far better than trying to recall benchmark scores.