GPT-6.1 Sol Service Incident Report

GPT-6.1 Sol Performance Collapse: Questions Surround OpenAI's Resource Allocation

Over the past six hours, multiple reports of anomalies concerning GPT-6.1 Sol have appeared on X, pointing to significant degradation in the model's response speed and execution efficiency. User reports indicate that routine tasks which previously took 10-20 minutes now generally require several hours or longer, with some scenarios approaching unusable status. This phenomenon aligns closely with YZ Index smoke test results, raising industry concerns about OpenAI's internal resource allocation and model stability management.

User feedback centers on two core issues: a cliff-edge drop in speed and diminished execution capability. One long-term user noted that wait times after submitting simple tasks surged from a few minutes to nearly 50 minutes, in stark contrast to Claude Fable 5.2's processing efficiency, suspecting a systemic bottleneck in OpenAI's internal processes. Another English-language report directly questioned "whether I'm the only one experiencing severe slowdowns with GPT-6.1 Sol," speculating that OpenAI is reserving GPU resources for GPT-6.2, thereby squeezing existing services. A third user mentioned that "GPT-6.1 Sol has been weakened," with routine project runtimes substantially extended, suspecting the model has entered a "dark path." Though phrased differently, all three reports point to the same fact: the model's response latency and output quality are declining in tandem in real-world engineering scenarios.

YZ Index smoke tests provide further quantitative evidence. The model's main leaderboard score plummeted from 84.4 to 67.71, a drop of 16.69 points. A targeted retest using an enhanced 18-question version showed an execution capability score of 79.50 and a robustness score of only 53.30, with the latter showing a particularly dramatic decline. The robustness metric directly reflects the model's consistency and interference resistance in long-chain tasks; a low score here means GPT-6.1 Sol is highly prone to losing intermediate states or output degradation when handling multi-step engineering tasks, fully consistent with users' intuitive sense of "tasks getting stuck."

From a technical perspective, the most likely trigger for this anomaly is compute reallocation. OpenAI is in a critical window for GPT-6.2 training, and historical patterns show that whenever a new model enters large-scale pre-training, GPU quotas for existing service clusters are often dynamically reduced to ensure training efficiency. If GPT-6.1 Sol's inference instances are batch-migrated to low-priority queues or concurrency quotas are restricted, this would directly manifest as the "extremely slow" experience users perceive. Second, the model itself may have undergone insufficiently validated weight updates or alignment fine-tuning, causing its original capability boundaries to contract. The sharp drop in robustness scores suggests possible "catastrophic forgetting" or degraded instruction-following, which is not uncommon in rapidly iterating closed-source models. Additionally, at the infrastructure level, hidden failures in inference clusters or imbalanced load-balancing strategies could produce similar symptoms.

Regardless of whether the root cause points to resource contention or model degradation, this incident exposes shortcomings in OpenAI's service stability and transparency. Users' dependence on a single model has far exceeded previous levels, and any unexpected performance fluctuation directly impacts productivity. It is recommended that OpenAI promptly disclose GPT-6.1 Sol's current quota policies and rollback plans, while strengthening continuous benchmark monitoring for engineering scenarios to prevent similar "dark path" incidents from recurring. For developers relying on this model, switching to multi-model parallel validation or localized solutions in the short term may prove a more prudent response strategy.