This is absolute chaos...

2026-07-13 · Source: Theo - t3․gg · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering · Depth: Intermediate, extended

Summary

GPT56 Soul, while powerful, significantly increases token usage compared to GPT55, leading to rapid exhaustion of rate limits. OpenAI has temporarily removed the 5-hour usage limits for Plus, Business, and Pro plans, leaving only the weekly limit, and is rolling out efficiency changes for GPT56 Soul. Key strategies to mitigate high usage include avoiding the "Ultra" feature and "Fast mode," which burns limits 2.5x faster. Benchmarks like Deep SWE show "High" reasoning level offers a 69% score at \$3.47 per task, providing a strong balance of performance and cost, making "X high" and "Max" less efficient for marginal gains. The model's eagerness to continue requires users to explicitly define stop conditions within prompts. Additionally, managing sub-agent behavior and avoiding manual context window limits (which OpenAI confirms are optimally tuned at 272k/372k) are crucial for efficient operation. The author encourages experimentation with settings and prompts rather than blindly following external advice.

Key takeaway

For AI Engineers optimizing GPT56 usage, you must actively manage its eager execution to control costs and avoid rate limits. Prioritize "High" or "Medium" reasoning levels and disable "Fast mode" to significantly reduce token burn. Crucially, embed explicit stop conditions within your prompts to guide the model's output length. Additionally, consider adjusting your agents.md to restrict sub-agent use, preventing unintended token consumption. Experiment with these settings to tailor the model's behavior to your specific workflow and budget.

Key insights

GPT56's high usage necessitates strategic prompting and configuration to manage costs and rate limits effectively.

Principles

• Explicitly define model stop conditions in prompts.
• Higher reasoning levels often yield diminishing returns.
• Avoid "Fast mode" for significant usage reduction.

Method

Configure global agents.md to restrict sub-agent use unless explicitly requested. Set reasoning level to "High" or "Medium" for optimal cost-performance. Craft prompts with clear stopping points to control token generation.

In practice

• Turn off "Fast mode" in GPT56 settings.
• Set default reasoning to "High" for most tasks.
• Add "Only use subagents if the user explicitly requests them" to agents.md.

Topics

• GPT56 Soul
• Rate Limits
• Token Usage Optimization
• Prompt Engineering
• Reasoning Levels
• Sub-Agents
• OpenAI Codex

Best for: AI Engineer, Software Engineer, Prompt Engineer

Related on AIssential

• GPT-5.6: The Review — GPT-5.6 Soul sets new intelligence and efficiency standards, excelling in complex tasks with advanced agentic capabilities.
• I Cut My AI Coding Assistant’s Token Usage in Half. Here’s Exactly How — Inefficient context management, not subscription limits, drives high AI coding assistant token usage.
• Semantic Early-Stopping for Iterative LLM Agent Loops — Semantic early-stopping for LLM agent loops reduces token cost by halting when meaning or quality plateaus, outperforming fixed iteration caps.
• Getting more from each token: How Copilot improves context handling and model routing — GitHub Copilot now intelligently routes tasks to optimal models and caches context for enhanced efficiency.
• Claude Code Was Burning Tokens Until I Put a Gate in Front of It — Enforce AI agent behavior at the system level with auto-correcting hooks to prevent context bloat and optimize token usage.
• You aren't using Codex like me... — Optimizing Codex with GPT-5.6 requires strategic model selection, advanced features, and robust safety configurations for peak performance and security.

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.