Every week our pipeline scrapes the model catalogs and vendor blogs, then judges each item against one question: would this actually improve something we run in production? Verdicts below. Watch = interesting but unproven claims · Adopt = earned a place · Ignore = noise · Product Input = a feature idea, not a factory change.
W28 review: 30 items across LlamaIndex blog (15) and model-watch (15). Heavy deduplication applied — 7 items dropped as duplicates across two LlamaIndex clusters (LiteParse ×4→1, agentic OCR ×4→1, retrieval rerank ×2→1). All surviving LlamaIndex items land at Watch: vendor blog source scores 7, below the 8/10 promotion threshold; mechanisms are sound but require primary-source verification before adoption. No Factory Candidates or Immediate Upgrades this week. Model-watch highlights: Sonnet 4.6 gains a new optional 1h cache-write tier — incumbent across 10 apps, additive change, worth cost-modeling for burst-heavy apps. Two Opus 4.8 variants (base at $5/$25/M, fast at $10/$50/M OpenRouter) confirmed real models; both Watch pending eval scorecards. Two Google image models (Gemini 3.1 Flash Image, Gemini 3 Pro Image) relevant to FloodStream/FPA — Watch pending eval, specific specs flagged for verification. Two deprecations (gpt-4o-audio-preview, deepseek-v3.2-speciale) need one grep each to confirm non-incumbency. Six model-watch items dropped as Ignore (negligible price deltas, confirmed non-incumbents, untrusted publisher). No backlog rows created this week.
New Anthropic Opus 4.8 base model via OpenRouter; 1M context, trusted publisher, half the cost of the fast variant. Highest-quality tier candidate for complex reasoning, arbitration, and judge tasks.
Why this verdict: Real model confirmed by session system context (Opus 4.8 listed as current). OpenRouter pricing ($5/$25/M) needs independent verification against Anthropic direct pricing. Watch until eval scorecard arrives; strong candidate for judge-tier tasks if eval holds.
Faster variant of Opus 4.8 at higher cost ($10/$50/M input/output) but lower latency; 1M context. Targets latency-sensitive tasks where Sonnet quality is insufficient but standard Opus latency is blocking.
Why this verdict: Real model family (Opus 4.8 confirmed). Fast variant is premium-priced ($10/$50/M OpenRouter) and only justified if latency improvement is material. Watch until scorecard; likely narrower use case than base Opus 4.8. OpenRouter pricing needs verification.
Image-specialized Google model (Gemini 3.1 Flash Image) at low cost via OpenRouter; optimized for vision tasks. Internal name 'Nano Banana 2' suggests image-first architecture. Potential fit for photo-heavy workflows.
Why this verdict: Image-specialized model from a trusted publisher, auto-eval queued. Relevant to FloodStream and FPA but specific capabilities and pricing need verification before routing real photos through it.
Pro-tier image+audio+reasoning model (Gemini 3 Pro Image) at $2/$12/M with internal_reasoning pricing; 65K context. Stronger image analysis than flash variant at 4× the cost.
Why this verdict: Pro tier has smaller context window (65K vs 131K for flash) despite higher cost, which limits document-length inputs. Only justified if accuracy delta over flash is confirmed. Subordinate to in evaluation order.
Free-tier Kimi K2.6 model with 262K context on OpenRouter; zero prompt/completion cost enables cost-free baseline evaluation and development testing without consuming eval budget.
Why this verdict: Free tier models are useful for development iteration without burning eval budget. MoonshotAI is a credible lab but kimi-k2.6 is not yet evaluated in the Factory. Watch until scorecard arrives.