Optimizing GPT-5.6 Usage and Avoiding Rate Limits
Theo - t3.gggo watch the original →
the gist
GPT-5.6 is highly capable but prone to aggressive token consumption; users can maintain productivity by disabling 'Fast Mode,' avoiding 'Ultra' features, and capping reasoning levels at 'High' to maintain an optimal cost-to-performance ratio.
The Shift in Model Behavior
GPT-5.6 (specifically the 'Soul' model) represents a significant departure from its predecessor, GPT-5.5. While 5.5 was often constrained by its tendency to stop and request user feedback, 5.6 is designed to run autonomously for long durations. This shift in behavior—while powerful—has led to rapid depletion of rate limits. Users accustomed to the 5.5 workflow, where frequent interaction was required, are finding that 5.6 consumes tokens at a much higher rate per message, particularly when 'Fast Mode' is enabled.
Strategic Configuration for Efficiency
To mitigate limit exhaustion, users should adopt a more conservative configuration. First, 'Fast Mode' should be disabled; the speed gains are marginal compared to the 2.5x increase in token consumption. Second, the 'Ultra' feature should be avoided until further notice, as it is a primary driver of excessive usage. Finally, reasoning levels should be capped. Benchmarking data indicates that 'High' is the sweet spot for performance; moving to 'X-High' or 'Max' yields diminishing returns in code quality while significantly increasing token costs.
Managing Sub-Agents and Context
GPT-5.6 is prone to over-eagerly spinning up sub-agents for tasks that do not require them. To prevent this, users should update their agents.md file with a directive to only use sub-agents when explicitly requested. This simple constraint prevents unnecessary token burn. Furthermore, while the model is capable of long-running threads, users should remain mindful of context management to ensure the model doesn't fixate on irrelevant historical data.
The Future of Usage Limits
OpenAI has acknowledged the current friction and is actively rolling out efficiency improvements. As of the latest updates, the 5-hour rolling limit has been temporarily suspended in favor of a weekly limit. While this provides more flexibility for bursts of work, it requires users to be more disciplined, as a single poorly configured prompt could theoretically consume a significant portion of the weekly quota.