# You Ask, I Answer - Multi Agent AI Orchestration Costs > [!question] Lingering Questions > - What is the actual measured cost delta between model tiers on a documentation task? Penn asserts large savings but gives no figures. > - Is "writing code is relatively inexpensive" still true for agentic work on large codebases, where context recycling is also heavy? > - Where exactly is the boundary between organising and creating? Summarisation, translation, and refactoring all sit ambiguously. > [!tip] Article Ideas > - A short piece pairing the "organising vs. creating" test with the vault's existing credit-routing policy — check whether the routing matrix already covers it or should be extended. > [!link] Cross Reference > - [[3. Resources/Permanent Notes/Route Judgment to the Expensive Model and Bulk Execution to the Cheap One]] — same routing logic, framed by task type instead of by decision-vs-execution split > - [[3. Resources/MOCs/AI Cost & Tokenomics MOC]] > [!quote] Vocabulary > - **Heavy context penalty** — Penn's coined term for why documentation/consolidation tasks cost more per output token than code generation: the model must hold and re-read large text volumes in working memory. ## Bib capsule Christopher S. Penn answers a viewer question on his "You Ask, I Answer" YouTube series (2026-07-29) about why five readme files cost over $100 in Claude credits, and whether multi-agent orchestration can ever be economical. ## Author's thesis A $100 bill for five readme files is a routing failure, not evidence that orchestration is uneconomical. Documentation work carries a heavy context penalty — it holds and recycles large text in working memory — so it should run on the lightest model that can do it, and often on no model at all, since indexing has deterministic solutions. The deciding test: are you organising existing data or creating something new? ## Key arguments (≤3) 1. **The diagnosis is task-model mismatch, not orchestration overhead.** The viewer used the most expensive model and vendor for a mechanical task; that's a billing and routing error, not proof orchestration doesn't scale. 2. **The mechanism explains the cost.** Documentation is expensive because it requires holding and re-reading large files in working memory; cost tracks context recycled, not output length. 3. **The prescription is an operational test.** Organising data you already have doesn't need a reasoning model; creating something net new might. Apply the test before running the task, not after the bill arrives. ## Claim atoms ### Documentation work carries a heavy context penalty > [!quote] ID:rw_01kyq6v55ht821zday038b3bjx > It consumes a lot of working memory to read and write big text files. That's why writing code is relatively inexpensive. Defines the mechanism, not just the advice — explains why cost tracks context recycled rather than output produced, which transfers past the specific task Penn names. ### Multi-agent orchestration needs lightweight tuned models > [!quote] ID:rw_01kyq6veaab38ysj1fy97pxsre > My suggestion for multi-agent orchestration is use a lightweight cheap model, ideally on something like a token plan, that is tuned and optimized for those kinds of tasks. Direct confirmation of the credit-routing policy already in use: match tier to task, keep cheap models available for price comparison. ### Reorganising existing data needs no reasoning model > [!quote] ID:rw_01kyq6vxyswd7vemwfx1wgkdvg > if you were asking it to organize existing data, you're not asking it to make anything net new, use a light, fast model An operational test applicable before running a task — organise vs. create — which is what separates it from post-hoc cost analysis. ## Permanent candidates - **Inference cost tracks context recycled, not output produced** — the strongest, least-covered claim in the vault; promoted below. - The organise-vs-create routing test is a near-duplicate of the vault's existing [[3. Resources/Permanent Notes/Route Judgment to the Expensive Model and Bulk Execution to the Cheap One]] — connect rather than duplicate. ## Processing checklist - [x] Ported from existing deep-read (Refactor Appendix) - [x] Checked for duplicate Permanent notes — found one near-duplicate, linked instead of recreated - [x] Connected into MOC — [[3. Resources/MOCs/AI Cost & Tokenomics MOC]] - [ ] Find the referenced May 2026 Almost Timely newsletter issue on "18 ways to save token budget" and extract the list ## Metadata - Source: [[3. Resources/Readwise/Articles/You Ask, I Answer Multi Agent AI Orchestration Costs|You Ask, I Answer - Multi Agent AI Orchestration Costs]] ## Links - [[3. Resources/MOCs/AI Cost & Tokenomics MOC]] — routing-by-task-fit is exactly this MOC's organizing question - [[3. Resources/Permanent Notes/Route Judgment to the Expensive Model and Bulk Execution to the Cheap One]] — same cost-routing logic from a decision/execution split rather than an organise/create split