Ultra Mode Is a Budget Burner, Not a Reasoning Level

OpenAI launched 'Ultra' as a reasoning option in Codex, but it's actually a system prompt instruction to spawn sub-agents at max reasoning effort. This misclassification leads to rapid rate-limit exhaustion—the speaker burned a 5-hour limit in 20 minutes. Teams expecting improved reasoning are instead handed an ungoverned cost accelerant.

Codex’s V2 architecture shares full conversation history with all child agents, permits unlimited recursion with no depth limit, and locks all sub-agents at max reasoning. This multiplies token usage far beyond earlier versions, and the unfinished default now makes new models a financial risk.

Claude Code's 'Ultra Code' reveals a safer pattern: it uses programmatic workflows with fixed phases, schemas, and controlled reasoning (defaulting to X-high, not max). This gives multi-agent power with predictable endpoints. OpenAI copied the aggressive agent trigger but omitted the guardrails, resulting in runaway costs instead of productivity.

Engineering leaders should block Ultra usage, treat AI tool spend as a budget-governed resource, and evaluate alternatives like Claude Code’s workflow-based agents. The broader signal is that multi-agent features are shipping without adequate governance, turning tool adoption into an infrastructure-cost decision.