First screen, captured 7 days ago. Open the video.

A Jev that can ask Opus for help gives Opus accuracy at 4x lower cost

What HN thinks

Everyone has a harness, nobody agrees which model should be the boss, and a side quest broke out over whether '4x lower' is legal arithmetic.

Mix cheap and smart models
Many commenters already run tiered setups, either a strong model orchestrating cheap workers or a cheap/local model escalating to Opus, and report big cost savings. Which direction works best seems task-dependent.
Escalation hinges on knowing uncertainty
The hard part is a weak model recognizing when it needs help; calibrated confidence makes it work here, and the threshold is a cost-vs-correctness knob.
Narrow result, worrying application
Skeptics note the benchmark is only identity resolution, question the calibration analysis and data ambiguities, and some find AI-assisted record matching a privacy concern.

Consensus: Routing between cheap and expensive models can cut costs substantially, but results are task- and model-version-specific rather than a general rule.

Recommended in the comments

OpenCode ↗ (tool): Lets you assign different models to orchestrator and subagent roles without being tied to one vendor.
Claude Code advisor ↗ (product): Native way to have Haiku or Sonnet escalate to Opus when needed.
llm (tool): Simple way to let Opus prompt a local Qwen model.
JevK5 ↗ (library): One commenter found it nearly matched Jev on their workload locally.

Summarised by Claude from the comment thread 6 days ago.

Points: 568
Comments: 142
Best rank: #1
Posted: 2026-11-06 by theanomaly
Category: AI & ML

identity-resolution llm-as-judge claude-opus decision-models model-cascading hightouch benchmarks

All captures:

Full page

Captured 7 days ago. Click for the whole screenshot.

Looks similar

Screenshot of resobscura.substack.com: Using LLMs to trace alchemical knowledge and decode 17th century letters

Using LLMs to trace alchemical knowledge and decode 17th century letters (resobscura.substack.com)

160 points by benbreen 1 month ago | 38 comments

A Res Obscura Substack post by historian Benjamin Breen arguing AI labs should fund historical research, with examples of using frontier LLMs to trace alchemy and decode 17th-century letters.

llms historical-research alchemy digital-humanities

Screenshot of johnwhiles.com: AI slows down open source developers. Peter Naur can teach us why

AI slows down open source developers. Peter Naur can teach us why (johnwhiles.com)

303 points by jwhiles 1 year ago | 187 comments

Screenshot of blog.alexellis.io: Local Qwen isn't a worse Opus, it's a different tool

Local Qwen isn't a worse Opus, it's a different tool (blog.alexellis.io)

433 points by alphabettsy 4 months ago | 227 comments

Blog post by Alex Ellis comparing local Qwen models with Claude Opus, arguing local models are a different tool rather than an Opus replacement, based on real project use.

qwen claude-opus local-llm ai-agents

Screenshot of david.lequin.dev: Can we replace a $150k Looker with a bucket of HTML files?

Can we replace a $150k Looker with a bucket of HTML files?(david.lequin.dev)

113 points by skadamat 21 days ago | 73 comments

Blog post by a data platform engineer describing how his team replaced much Looker usage with an internal service hosting Claude-generated HTML dashboards backed by Snowflake.

business-intelligence looker snowflake claude-code