The Ultra Mismatch

• "Ultra" is presented by OpenAI as a highest-reasoning toggle, but it functions as a skill-trigger that shifts prompting and activates broad subagent orchestration.5:45
• In one Claude Code implementation, selecting "Ultra Code" programmatically defaults the effort level to "XHigh" while appending workflow logic, contradicting the label of an independent, maximal-reasoning mode.
• The speaker notes that OpenAI recently hid the Ultra and "max" settings from certain UI sliders, suggesting the company is actively iterating on the labeling of these autonomous features.4:08

Implementation Risks

• Codex V2, the subagent system currently routing automatically for new models, forces full-thread history sharing by default, which creates massive token bloat and context noise.14:16
• The architecture lacks a subagent recursion depth limit, allowing models to fan out or recurse until agent termination conditions are met—a behavior that triggered the speaker to hit their 5-hour rate limit twice within sixty minutes.16:28
• Enabling V2 can cause exceptions when custom V1-based instructions or rate limits are present, as the two systems are currently in a transition state that lacks seamless compatibility.12:56

Workflow Alternatives

• Anthropic’s "Claude Code" workflow model is highlighted as a more mature architecture because it uses phase-based, bounded JavaScript configurations rather than open-ended recursive agents.19:58
• These phase-based workflows provide explicit stopping points and schema-defined handoffs, which arguably prevent the runaway token costs currently observed in the Ultra-enabled OpenAI setup.21:33