Voice-agent builders begged OpenAI not to kill gpt-realtime. Support answered Monday: no promises.
Two production voice-agent operators posted benchmarks showing the recommended replacements failing most of their calls. Seven weeks later the only official reply is that the regressions will be passed along. Then the thread was closed.
On July 20 OpenAI added four Realtime models to its deprecations page, with a shutdown date of January 20, 2027 and a recommended replacement for each. The original gpt-realtime and the older gpt-4o-realtime are to be replaced by gpt-realtime-2.1; their mini variants by gpt-realtime-2.1-mini. The next day a company that sells voice software to businesses opened a thread on OpenAI's developer forum with a title that is itself a plea: "Gpt-realtime now shows 'Deprecated'. Impact on deterministic voice agent architectures. Please do not deprecate."
What makes the thread worth reading is that it contains numbers rather than feelings. The poster, writing as WebPlanning_SIM-Ltd, runs a production voice service on the August 2025 snapshot of gpt-realtime and reports "98-100% success" with "~4s-6s response time." Against the newer generation the same system collapses. gpt-realtime-2, by the poster's measurement, manages "0-20% success, up to 14s response time." The 1.5 release sat around 75 percent and the 2.1 release, the one OpenAI recommends, did worse still in their testing. The complaint is architectural, not cosmetic: the newer models are reasoning-based, and a scheduling system that needs a model to follow a structured command exactly, every time, cannot tolerate one that improvises.
A second operator, the same failure
Two weeks later a second developer, posting as carlos_ds, described an attempted migration of outbound voice agents from the legacy mini model to gpt-realtime-2.1-mini. The list of failures is specific: "heavy conversation latency, structural hallucinations, severe instruction narration." The model, they wrote, narrates its own internal state transitions out loud to the person on the phone, and it over-interprets example dialogue in the prompt so literally that it skips real steps in the conversation. They asked for either an extension of the deprecation date or a switch to turn the commentary off.
Then nothing, for over a month. On Monday afternoon an account named OpenAI_Support replied. "The current notice lists January 20, 2027 for GPT-Realtime's shutdown," it wrote. "We hear that the suggested replacements haven't met your structured-tool requirements. We'll pass along those regressions and the retention/control requests, but can't promise an extension or fix timeline." The thread was closed shortly afterward.
What that answer means for the category
Voice is where a great deal of agent revenue actually lives right now. Outbound sales calls, appointment booking, tier-one support and the phone tree replacements that every mid-sized business is being pitched all run on realtime speech models, and OpenAI's is the one most of the wrapper companies chose. We wrote earlier this year that voice is becoming the front end for agents, and investors agreed. The operators in this thread are precisely that layer: small firms that took OpenAI's model, wrapped a deterministic control system around it and sold the result to businesses that now depend on it.
Their problem is not that OpenAI is retiring a model. Every provider does that, and a six-month notice period is standard. Their problem is that the officially recommended successor does not do the job, by their measurement, and that the company's answer on the record is that it cannot commit to fixing it before the old one is switched off. That leaves a builder with four months to either re-engineer around a model that narrates its own state machine, move to a rival provider, or hope. It is fair to note that only two operators posted benchmarks and that prompt design accounts for some share of any migration gap. It is also fair to note that OpenAI did not dispute a single figure.
There is a larger pattern here that GaaS buyers should recognize. Platform vendors are pushing every product line toward reasoning models because those are the models that make multi-agent orchestration work and that show well in demos. A reasoning model is a worse fit for the boring, high-volume, must-not-deviate tasks that make up most deployed voice agents. When the vendor's roadmap and the customer's workload diverge, the deprecation calendar decides who wins. OpenAI says its agent products reach 10 million weekly users. Some unknown fraction of them are going to get a very different phone call in January.
We will update this story if OpenAI extends the date or ships a fix for the instruction-following regressions its own support team has now acknowledged in writing.
