% open deep-research-lean
Deep Research, Half the Cost
Forked an expensive research tool to run for well under half the cost, same rigour.
PROBLEM
I lean on a deep-research tool most days, it fans out across the web, checks its own claims, and writes a cited report. The catch is the cost: one run fires off close to a hundred sub-agents, and most of that goes on grunt work, fetching pages and fact-checking, not the actual thinking. I was paying top-model prices for work a cheaper model does just as well.
WHAT I BUILT
A leaner version of the same research harness. Same shape, same rigour, designed to cost well under half. The trick is matching the model to the job: the smart, expensive model handles the parts that shape quality, a cheaper, faster one handles the mechanical parts.
HOW IT WORKS
A run has four stages: scope the question, fetch the sources, verify the claims, synthesise the report. Scope and synthesise are where the thinking lives, so they stay on the top model. Fetch and verify are mechanical, so they moved to a cheaper model at low effort. I also trimmed how many sources it pulls and how many claims it checks, but left the part that matters most untouched: every claim still gets challenged by three independent adversarial votes.
WHERE IT'S AT
Live and now my default for research. The saving is a design estimate from the changes, not a benchmarked figure yet, but a run should cost well under half the original. If one ever comes back thin, I fall back to the full version for max breadth.
WHAT I LEARNED
Most of the cost in an AI system isn't in the thinking, it's in the grunt work around it. Routing the mechanical stages to a cheap model and keeping the expensive one for scope and synthesis cut the bill hard without touching the rigour. The fact-check was the one thing I refused to weaken.
STACK
Claude Code workflow orchestration · multi-agent fan-out · mixed-model routing
% cd ..