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.
30 items judged. 25 dropped as recurring static blog-canon posts from eugene-yan/hamel-husain/lilian-weng (evals/agents/prompting writeups already known or already_backlogged, no new fact this week). 3 model-watch incumbent alias-repoint items promoted: sonnet-4.6 and haiku-4.5 are confirmed INCUMBENT:YES on 10 and 7 apps respectively (Immediate Factory Upgrade, outage-prevention exception, verify resolved snapshot against each app's regression baseline); opus-4.7 is INCUMBENT (UNVERIFIED) on 5 apps (Factory Candidate, first step is to confirm the pin then migrate to claude-opus-5 per standing Opus deprecation policy). 2 latent-space/AINews Tier-3 items (Claude Tag Slack agents; GLM-5.2 cost-swap claim) held as Watch pending primary-source verification of their specific claims.
AINews newsletter describes a 'Claude Tag' multiplayer/async/proactive agent layer in Slack with git webhooks and multi-agent handoffs, reportedly in beta for Team/Enterprise.
Why this verdict: Tier 3 (newsletter roundup, no primary-source link). Novelty Medium (new named feature) but materiality is generic — no named product or decision changes today. Held behind verification per the Tier-3 rule; Watch is appropriate until a concrete use case exists.
AINews digest item (mistitled after SpaceX, actual content is GLM-5.2) claims GLM-5.2 is now usable in agent harnesses (Cline, Baseten, Claude Code proxy) as a cheap swap for Sonnet-tier tasks.
Why this verdict: Tier 3 (newsletter roundup, title/content mismatch itself is a quality flag). Novelty Low-Medium (GLM family already known; this is an incremental model version). Per model-intake rule, a new model is only material if it plausibly lands on the efficient frontier, which is unverified here. Held behind verification; Watch, not Product Input.