Article illustration 1

OpenAI is racing to address mounting criticism of GPT-5’s inconsistent performance, with CEO Sam Altman publicly attributing key failures to a malfunctioning "autoswitcher" that hampered the model’s reasoning capabilities. The concessions follow BleepingComputer tests revealing stark disparities between GPT-5’s impressive coding speed and its struggles with creative tasks and detailed responses.

The Autoswitcher Breakdown

Altman confirmed a critical bug in the system designed to activate GPT-5’s advanced reasoning modules: "Yesterday, the autoswitcher broke and was out of commission... GPT-5 seemed way dumber." This admission explains widespread user reports of the model producing unusually truncated outputs and failing to engage its full cognitive potential—issues some initially suspected were cost-saving throttling measures.

Patch Rollout and User Controls

Key fixes now in motion include:
- Doubled rate limits for GPT Plus subscribers within days
- New UI toggle to force reasoning capabilities manually
- Clearer indicators showing which model processes queries
- Legacy model access: Plus users can revert to GPT-4o, though OpenAI warns this may be temporary

Performance Paradox

BleepingComputer’s analysis highlighted GPT-5’s paradoxical nature: blazing-fast code generation overshadowed by creative shortcomings. When explicitly prompted for elaboration, the model frequently delivered sparse responses—a flaw undermining its value for content-heavy workflows. Altman promises intelligence upgrades will roll out imminently, but the phased global deployment means many users won’t see improvements immediately.

Developer Impact

For engineers building atop OpenAI’s APIs, the turbulence underscores the risks of dependency on rapidly evolving proprietary models. The temporary restoration of GPT-4o offers a lifeline, but Altman’s caveat—"we'll watch usage as we think about how long to offer legacy models"—signals uncertainty. Meanwhile, the doubled rate limits ($20/month tier) alleviate compute constraints throttling data-intensive applications.

As OpenAI stabilizes GPT-5, the episode reveals the fragile equilibrium between breakthrough velocity and production reliability. For developers, the new manual controls offer tactical relief, but long-term trust hinges on consistent delivery of the "state-of-the-art" performance OpenAI promises.

Source: BleepingComputer