The detail that matters in this Quesma write-up isn’t that a deep-research run burned a whole Claude Max plan in 30 minutes. It’s where the budget went: the run launched 111 agents and queued 123 claims for verification, but only 25 got verified before the limit hit — and the final synthesis never ran. That’s the real shape of these systems. The search is cheap; the checking is what bankrupts you.

Once you internalize that, the architecture the author landed on stops looking clever and starts looking necessary. He split the work by role instead of throwing one expensive model at everything: Sonnet 5 to find (cheap enough to run in bulk), Opus 4.8 to verify (accuracy matters more on the check than the search), the priciest model reserved only to plan and resolve disputes, Haiku for extraction and formatting. That’s routing driven by where correctness actually needs to live, which is exactly the calculus I keep hitting in production agentic work — you don’t pay for the smartest model per token, you pay for it per decision that would be expensive to get wrong.

The part I hadn’t seen before: he wired Codex and Antigravity in as headless subagents under a Claude Code harness, sharing memory via an extended claude-mem, so tokens draw from three separate subscription pools he already pays for. Cross-vendor subagents as a rate-limit arbitrage. It’s a hack, but it’s the honest kind — a direct response to the fact that a single plan’s allowance can’t sustain a verification-heavy pipeline.

The HN thread is arguing about whether this counts as gaming the plans. I’d rather ask the design question: if verification is the cost center, should deep-research harnesses be quoting you a confidence budget up front, the way a query planner estimates rows before it scans?