Why ChatGPT Plus Performance Is Dropping and What OpenAI’s Roadmap Reveals

Users of ChatGPT Plus have recently reported a significant drop in platform performance, including a 13% decrease in programming accuracy, which they attribute to the underlying GPT‑4 model.

These issues emerged after a series of updates that added web‑browsing and extended plugin access for Plus subscribers, dramatically increasing the service’s workload and exposing GPU bottlenecks that slowed response times.

Faced with reduced functionality and speed, many users are considering canceling their subscriptions and turning to open‑source LLMs as an alternative.

The community has offered several explanations: some argue that the performance drop stems from “model scaling pain,” where reducing GPT‑4’s inference capacity is a plausible way to balance speed, while others point to extensive AI‑alignment efforts that sacrifice accuracy for safety.

Research from Microsoft Research, cited by Sebastien Bubeck, supports the claim that AI alignment can degrade model performance; an unaligned GPT‑4 can generate high‑quality images from a TikZ prompt, whereas the aligned ChatGPT version produces noticeably poorer results.

In a recent discussion with Humanloop’s CEO Raza, Sam Altman outlined OpenAI’s short‑term roadmap, emphasizing that GPU scarcity is delaying many initiatives. Key priorities include delivering a cheaper, faster GPT‑4, extending context windows up to 1 million tokens, expanding the finetuning API, and introducing a stateful API that remembers conversation history.

Other announced items are multimodal capabilities (still limited by GPU availability), a dedicated‑capacity offering that costs $100 k upfront, and a cautious stance on releasing ChatGPT plugins, which Altman believes lack product‑market fit.

Altman also stressed the need for regulation of future models while advocating for open‑source releases, noting that OpenAI is considering open‑sourcing GPT‑3 but remains wary of the infrastructure required to host large LLMs.

Despite claims that the era of ever‑larger AI models is ending, internal OpenAI data suggest that scaling laws still hold, implying continued performance gains from larger models and a potentially shorter timeline for achieving AGI.