GPT-6.1 Sol Service Anomaly Report

GPT-6.1 Sol Performance Cliff: Service Anomaly Under 46 User Reports

Over the past six hours, 46 concentrated complaints about OpenAI’s GPT-6.1 Sol appeared on X, marking a clear anomaly in the model’s service. Multiple paying users pointed directly to a sharp drop in response speed and broken execution, raising questions about the value of premium subscriptions. The incident quickly gained traction, exposing the operational pressure on AI vendors after密集 new-model releases.

User feedback showed a high degree of consistency. A $200 subscriber said bluntly that “GPT-6.1 Sol is as slow as a turtle, while Astra burns through usage like a rabbit,” making the double waste of time and quota a core pain point. Another post pointed out the irony: several AI giants had previously called for “slowing down development,” yet launched at least nine new models in a short period, including Anthropic’s Claude Opus 5.5 and Sonnet 5.5, OpenAI’s GPT-6 Sol, GPT-6 Luna, and GPT-6.1 Sol, as well as Google’s Gemini 3.8 series. Codex subscribers also found in testing that under identical prompts, GPT-6.1 Sol’s tokens/sec was significantly below expectations, calling it a “deliberately throttled version.” These voices collectively point to a broken experience for paying users.

The YZ Index smoke evaluation provides quantitative support. The model’s main leaderboard score fell from 87.4 to 77.15, a drop of 10.25 points. A targeted enhanced 18-question retest further confirmed the anomaly: the execution score was only 78.00, and the solidity score was 76.10, both below historical averages. The evaluation team noted that the model showed significant degradation in multi-step reasoning and long-context consistency, closely matching the “slow + weak” phenomenon reported by users.

Analysis of possible causes points to three areas. First, the dense rollout of multiple models in a short period led to imbalanced compute scheduling; GPT-6.1 Sol may have been assigned lower-priority resources, causing a systemic decline in TPS (tokens per second). Second, rate-limiting policies for subscription tiers such as Codex appear to have been broadened, intended to control high-cost inference spending but directly harming user perception. Third, rapid iteration of version 6.1 may have introduced insufficiently validated inference optimizations, causing both execution and solidity to decline. Whichever factor is responsible, it reflects the tension between AI vendors’ “safety calls” and commercial expansion.

This anomaly puts pressure on brand trust in OpenAI. If paying users continue to experience a “double loss of time and quota,” they will accelerate their switch to competitors. At the industry level, stability control after dense releases has become a new bottleneck. OpenAI should disclose the root cause and fix timeline as soon as possible; otherwise similar incidents may become the norm.