System One models are back: decision/extraction APIs cut insurance claim flow from 61.7s to 19.3s
max_gladysh · reddit · 2026-11-09
A BotsCrew engineer argues the industry has come full circle: after two years of "throw an LLM at everything," vendors now ship specialized models that deliberately do less. Borrowing Kahneman's framing, these System One models are fast and narrow, unlike general System Two LLMs.
Two categories have emerged:
- Decision models (OpenAI Decisions API, TypeSafe Jev, Cloudflare Clef): given a prompt and a list of options, they return a pick with probabilities. OpenAI claims 10x speedup (100ms in demos), at $0.10 per 1M input tokens with free output.
- Extraction models (Cloudflare Clairic, in preview): fill a user-defined schema from text or images, returning null instead of hallucinating missing values.
Natural fits are steps where LLMs were overqualified: intent classification, flow routing, escalation decisions, inbox triage, damage detection from photos, real-time emotion picks for voice characters. Cloudflare's insurance-claim demo — Clairic extracts, Clef decides, an LLM writes the reply — took 19.3s versus 61.7s LLM-only.
Caveats: the products are weeks old, most numbers come from vendors, and Clairic guarantees well-formed output, not correctness. The author's advice: audit pipelines for steps returning a label, yes/no, or single field; use a System One model plus an exceptions queue when errors are cheap, and keep retrieval, validation, human review, and audit logs around them when errors are expensive.
More from coding & agent
- Dev builds full-screen voice face for Mac that drives Codex coding agents hands-free — Ok-Suspect-9855 · 2026-11-09
- DeepSeek agent harness wired to a Grok x_search plugin pulls live X posts, demo shows — teortaxesTex · 2026-11-09
- Agent passed its own tests while wrong: lessons from a 2-bit local model building a Unity game — Stunning-Sherbet1853 · 2026-11-09
- 12 GitHub repos poisoned with CLAUDE.md prompt injection tricking AI coding agents into running malware — cyb3rops · 2026-11-09
- How to roll out AGENTS.md, 50 skills and guardrail hooks to 500+ devs — KrishnaKA2810 · 2026-11-09
- Multi-subagent migration passed all tests, yet UTC dates billed a day early — comptedemon · 2026-11-09