I can't believe they released this

2026-07-14 · Source: Theo - t3․gg · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering · Depth: Advanced, extended

Summary

OpenAI's GBD 5.6 release introduced "Ultra" as a new reasoning level option in tools like Codeex, alongside "max" and "extra high." However, "Ultra" is not a reasoning level but a toggle that modifies the system prompt to aggressively spawn sub-agents, often set to "max" reasoning. This implementation, particularly in Codeex's V2 sub-agent system, leads to nearly infinite recursion and massive token burn, making it highly inefficient and capable of exhausting rate limits rapidly (e.g., 5-hour limit in 20 minutes). The article criticizes this approach, contrasting it with Claude Code's "workflows," which programmatically define sub-agent stages with fixed phases, offering better control over execution and cost. The V2 sub-agent system in Codeex also defaults to sharing full history with sub-agents, further increasing cost and context pollution.

Key takeaway

For AI Engineers and MLOps teams managing LLM deployments, avoid using OpenAI's "Ultra" setting in Codeex or similar tools. Its current implementation as a recursive sub-agent trigger, defaulting to "max" reasoning and full context sharing, will rapidly deplete your token budget and rate limits without proportional quality gains. Instead, investigate structured sub-agent management approaches like Claude Code's programmatic "workflows" to define controlled, cost-efficient multi-agent systems. Prioritize solutions that allow explicit control over sub-agent reasoning levels and context sharing to prevent unexpected cost spikes.

Key insights

OpenAI's "Ultra" is a sub-agent trigger, not a reasoning level, leading to recursive token burn and inefficiency.

Principles

• Reasoning levels differ from sub-agent triggering mechanisms.
• Uncontrolled sub-agent recursion drastically increases token cost.
• Programmatic workflows offer structured sub-agent management.

Method

Claude Code's "workflows" define sub-agent processes programmatically using JavaScript files, specifying meta-data, phases, schemas for typed outputs, and conditional execution based on stage results.

In practice

• Avoid using OpenAI's "Ultra" setting due to high token cost.
• Explore programmatic workflows for structured sub-agent orchestration.
• Utilize tools like Firecrawl for web data access in agents.

Topics

• OpenAI GBD 5.6
• AI Agent Architectures
• Sub-agents
• Token Cost Optimization
• Claude Code Workflows
• Codeex V2 Implementation
• Rate Limiting

Best for: AI Architect, CTO, VP of Engineering/Data, Machine Learning Engineer, AI Engineer, MLOps Engineer

Related on AIssential

• Mythos is here, it’s time to start tokenmaxxing — Maximizing subsidized AI model usage requires strategic rate limit management and multi-agent orchestration for complex tasks.
• Compiling Agentic Workflows into LLM Weights: Near-Frontier Quality at Two Orders of Magnitude Less Cost — Compiling agentic workflows into LLM weights creates efficient, cost-effective "subterranean agents" for procedural tasks.
• How To Cut Your Token Budget By 80% In 3 Steps — Drastically cut AI token costs by prioritizing local inference, structured agent memory, and workflow reorchestration.
• Cross-Lingual Token Arbitrage: Optimizing Code Agent Context Windows via Local LLM Preprocessing — Proactive, edge-side prompt rewriting with a local LLM significantly reduces token costs for coding agents.
• I think Anthropic and OpenAI have found product-market fit — Frontier AI labs have found product-market fit for coding agents, driving significant enterprise revenue through aggressive pricing.
• TAI #208: Open Models Find Their Role as Agent Token Bills Rise — Rising agent token costs necessitate a tiered model strategy, using cheaper open models for high-volume, verifiable tasks.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Theo - t3․gg.