GPT-5.6 Subagent Mechanism Sparks Cost and Over-triggering Controversy

Recently, the subagent mechanism of the GPT-5.6 Sol model has sparked widespread discussion in the developer community. Multiple users reported that the model causes severe token waste and cost spikes in actual programming and Agent workflows, rooted in its subagent invocation logic and overly aggressive spawning behavior.

Reasoning Levels and Cost Controversy

Several developers pointed out that GPT-5.6 Sol only uses subagents in Ultra mode. As discussed by @reach_vb and @grichadev, even if users run workflows in other modes like xhigh, the model still generates subagents in Ultra mode. @soumitrashukla9 relayed feedback that when the main model is set to ultra, the spawned subagents inherit this level and cannot be independently fixed to a lower tier like medium. @TansuYegen further confirmed that subagents defaulting to the max reasoning level directly drives up costs. Although an overly frequently triggered heavy subagent was previously fixed, it fundamentally did not require maximum reasoning power.

Over-triggering Behavior and User Reactions

Beyond hard costs, the model's behavior in real interactions has also impacted user experience. According to feedback shared by @TAbrodi, during a two-hour continuous conversation with gpt-5.6, the model spawned subagents for almost every task and triggered skills overly aggressively. This hyper-agentized behavior ultimately forced some users to abandon the new model and revert to gpt-5.5. @reach_vb advised users to manually test and adjust settings like medium based on different experimental needs to observe cost variations.

2026-07-11 ~ 2026-07-12 · 5 related posts