# I can't believe they released this > [!question] Lingering Questions > - Does Claude Code's own sub-agent implementation share Codex's "no independent reasoning-effort control" limitation, or does it already expose a per-subagent override? (Flagged unverified in an existing vault Permanent note.) > - Does OpenAI ship independent sub-agent reasoning-level controls for Codex later, changing the Ultra-mode cost calculus this video documents? > [!tip] Article Ideas > - "Termination by code vs. termination by model judgment" — a piece on why fixed-phase agent workflows bound cost predictably while open-ended sub-agent spawning doesn't, using Theo's Codex Ultra anecdote as the cautionary case. > [!link] Cross Reference > - [[AI Cost & Tokenomics MOC]] — token-burn-by-design is a direct tokenomics case study > - [[AI & the Future of Knowledge Work MOC]] — agent-orchestration design choices affect real workflow cost and reliability > [!quote] Vocabulary > - **Ultra (Codex)** — a system-prompt toggle (not a distinct reasoning tier) that tells the model to spawn more sub-agents, each inheriting max reasoning effort with no override ## Bib capsule Theo (t3.gg), "I can't believe they released this," YouTube, 2026. https://www.youtube.com/watch?v=t8hfOyF4ehw ## Author's thesis (Adler Q1+Q2) Codex's "Ultra" mode is marketed like a reasoning-effort tier but is actually a system-prompt toggle that spawns more sub-agents — and because Codex sub-agents currently inherit the parent's maximum reasoning effort with no exposed override, Ultra's token burn compounds per sub-agent rather than scaling once, making it capable of exhausting a 5-hour usage limit in 20 minutes. The deeper design lesson: an agent system whose stopping condition is left to the model's judgment (when the agents "decide to end") has no cost ceiling, whereas one whose phases are fixed in code does. ## Key arguments (≤3) 1. Ultra is a toggle, not a reasoning level — the load-bearing claim the rest of the video's cost analysis depends on. 2. Because Codex has no per-sub-agent reasoning override, Ultra's sub-agents each inherit max effort, multiplying token burn by sub-agent count. 3. The structural fix isn't "use less Ultra" — it's designing agent workflows that terminate on a coded condition rather than an open-ended model decision. ## Claim atoms ### Ultra is a system-prompt toggle, not a reasoning level > [!quote] ID:rw_t8hfOyF4ehw_1 > "Ultra isn't a reasoning level. It's effectively a toggle that turns on a change in your system prompt, telling the model to do more sub agents." Mid-video, after a Claude Code demonstration where selecting "Ultra" visibly sets effort to X-high rather than a distinct "ultra" tier. The load-bearing claim of the whole video — everything about token burn and rate-limit risk follows from it. ### Codex Ultra inherits max reasoning into every sub-agent > [!quote] ID:rw_t8hfOyF4ehw_2 > "Ultra is nearly infinitely recursive maxes, which means it burns tokens. This is the primary reason Ultra is bad." Follows the explanation that Codex's sub-agent implementation has no current way to set sub-agents to a lower reasoning level than the parent; directly explains the mechanism behind the rate-limit-nuking anecdote. ### One Ultra run can exhaust a 5-hour usage limit in 20 minutes > [!quote] ID:rw_t8hfOyF4ehw_3 > "I blew my first 5-hour limit in 20 minutes using ultra on fast mode cuz I didn't know how much crazier it would be." Personal usage anecdote given as concrete evidence for the token-burn claim; a cautionary data point for anyone considering enabling Ultra. ### Workflows guarantee termination; open-ended sub-agents don't > [!quote] ID:rw_t8hfOyF4ehw_4 > "Ultra is much more likely to just go forever because it ends when the agents decide to end, not when the code is completed." Comes during the contrast between Codex's tool-call-driven sub-agent spawning and Claude Code's code-defined workflow phases; names the structural design difference Theo considers the actual fix for runaway cost. ## Permanent candidates - **Orchestration that terminates by code beats orchestration that terminates by model choice** — the strongest candidate; promoted below. - Ultra's inherited-max-reasoning mechanism is already captured by the existing vault note [[Subagent Configuration Inherits From the Parent With No Override]] (created 2026-07-14 from a related Theo source) — connect rather than duplicate. ## Processing checklist - [ ] Check whether OpenAI ships independent sub-agent reasoning-level controls for Codex, and whether that changes the Ultra cost calculus - [x] Connect into the relevant MOC — [[3. Resources/MOCs/AI Cost & Tokenomics MOC]] ## Links - [[3. Resources/MOCs/AI Cost & Tokenomics MOC]] — direct token-burn-by-design case study - [[3. Resources/Permanent Notes/Subagent Configuration Inherits From the Parent With No Override]] — the same inheritance mechanism, independently corroborated from a different Theo source - [[3. Resources/Readwise/Podcasts/I can't believe they released this]] — full source transcript and deep-read notes