A viral technical finding explains why ChatGPT feels sluggish in extended sessions, and the culprit is a client-side design choice in GPT-4.5 Turbo, not OpenAI's infrastructure.
If you've noticed ChatGPT getting noticeably slower the deeper you get into a conversation, you're not imagining it, and OpenAI's servers aren't to blame. A technical disclosure that broke on Reddit Saturday and swept across X within hours points to something far more specific: a design behavior in GPT-4.5 Turbo that causes the model to re-process an increasingly bloated input window with every single prompt you send.
The mechanism at the center of this has been labeled "ghost tokens" by the user who surfaced it, identified on Reddit as 'context_window'. These are metadata artifacts retained in the local conversation state to allow quick regeneration of previous turns. The problem is they accumulate silently, and past roughly 35 turns or around 8,000 combined tokens, the overhead compounds into latency that looks and feels exactly like network lag. It isn't. The delay is happening before your request even meaningfully hits OpenAI's servers.
GPT-4.5 Turbo rolled out broadly in late March 2026, and the timing lines up with the uptick in user complaints about sluggish responses that followed. The ghost token behavior appears to be baked into how this architecture handles history context, prioritizing regeneration speed for earlier turns at the cost of processing efficiency as sessions extend. OpenAI has not officially commented on the mechanic, and neither Sam Altman nor other executives have addressed it publicly as of writing.
That silence matters, because the nature of the problem actually carries a relatively optimistic market read for OpenAI. A ghost token accumulation issue is a software optimization problem, not a hardware capacity crisis. Fixing it doesn't require spinning up additional GPU clusters or significant capital expenditure. It's a patch, not an infrastructure overhaul, which is a meaningful distinction for a company currently valued at a scale where every cost signal gets scrutinized.
The enterprise exposure is real
Where this gets more complicated is in enterprise deployments. Corporate clients running long-context agent chains, the kind used in multi-step document analysis, legal review workflows, or extended coding sessions, are effectively hitting this bottleneck in production. For those users, the performance degradation isn't a mild annoyance. It's a workflow liability, and the discovery arrives at a moment when AI procurement teams are under pressure to justify spend on productivity grounds.
Anthropic has been quick to benefit from the conversation, at least in terms of attention. Claude 4's fixed-state architecture, which the company has marketed as avoiding full history re-processing on each turn, is now being cited directly in threads discussing the OpenAI issue. Whether that translates into enterprise switching at meaningful scale remains to be seen, but the timing of the discourse is useful for Anthropic's positioning going into what has been a competitive mid-year sales cycle for enterprise AI contracts.
The broader takeaway from this episode is less about OpenAI specifically and more about where the AI efficiency conversation is heading. Parameter count and benchmark scores dominated the last two years of model marketing. What's becoming clear in 2026 is that inference efficiency, how gracefully a model handles real-world usage patterns rather than curated test prompts, is increasingly the metric that enterprise buyers actually care about. A model that performs beautifully at turn five and falls apart at turn forty isn't solving the workflows that matter most to paying customers.
For now, the practical workaround for affected users is straightforward: start a new conversation thread rather than extending an existing one past the point where slowdown becomes noticeable. It's not elegant, but it sidesteps the ghost token accumulation entirely. Longer term, watch for a quiet GPT-4.5 Turbo update that addresses the history context handling. If OpenAI ships one without much fanfare, that will confirm the fix was simpler than the viral discourse made it sound.
Also read: Google Courts Marvell for Custom AI Chips, Challenging Broadcom's Silicon Grip • llama.cpp merges speculative checkpointing and local AI inference takes a significant leap forward • AI's Hidden Bottleneck Is Not Silicon. It Is Copper.
