Gpt-realtime now shows "Deprecated" — impact on deterministic voice agent architectures. Please do not deprecate

API Deprecations
,
1

Hi there,

We run a B2B voice-driven planning SaaS (WebPlanning-AI) built on the Realtime API. Since November 2025, we’ve built an entire architecture around ‘gpt-realtime’ (based on alias snapshot ‘gpt-realtime-2025-08-28’) : a non-reasoning, with temperature management, snapshot that follows strict, structured tool-calling instructions with very low variance.

On our production test corpus (voice commands mapped to structured tool calls with conditional parameters), I measure:

  • gpt-realtime (gpt-realtime-2025-08-28): 98-100% success, ~4s-6s response time (max)

    gpt-realtime-1.5: ~75% success

  • gpt-realtime-2: 0-20% success, up to 14s response time (preambles + reasoning tokens add ~6s of overhead per tool call, and the model frequently ignores conditional parameter constraints defined in the tool description)

  • gpt-realtime-2.1: even worser

We noticed gpt-realtime (based on snapshot ‘gpt-realtime-2025-08-28’) is now flagged “Deprecated” in the model catalog, which was not mentioned month ago.

Questions for the community:

  1. Has anyone received an official deprecation notice for this specific snapshot?
  2. Is anyone else running production voice agents where the newer reasoning-based models (gpt-realtime 1.5, gpt-realtime 2 et gpt-realtime 2.1) measurably regress on strict instruction-following / structured tool calling?

This isn’t a complaint about new model quality in general — gpt-realtime-2.1 is clearly built for a different use case (conversational agents benefiting from reasoning and tool transparency).

But for deterministic domains (scheduling, resource planning, transactional voice commands), the newer (model gpt-realtime > 1.5) default behaviors (preambles, reasoning-driven parameter interpretation) break what used to be a reliable contract.

Our architecture depends on deterministic tool routing rather than autonomous reasoning.

The issue is not that the newer models are “worse”. They optimize for a different objective.

Hence upgrading from gpt-realtime to gpt-realtime-2.x changes the behavioral contract in a way that breaks production systems relying on deterministic instruction following.

We please urge you not to remove this model until you have a deterministic model that is at least as good as ‘gpt-realtime’

2

Production Migration Regression: gpt-realtime-mini → gpt-realtime-2.1-mini

Our team has spent the last couple of weeks attempting to migrate a production outbound voice agent (highly structured telephony workflow/volume calling) from the legacy realtime-mini snapshot to gpt-realtime-2.1-mini. Despite adhering strictly to the new prompting guidelines, we have encountered severe architectural regressions that prevent production deployment.

Core Regressions Observed:

  1. Instruction & Commentary Leakage: The model frequently narrates its internal state machine transitions aloud (e.g., “Let me continue with the call…”, “Bringing you into the identity step…”). Even with aggressive negative prompting to suppress commentary, the tight linkage between audio/text modalities causes internal reasoning tokens to leak into final user-facing audio.
  2. Literal Transition Over-Matching: 2.1 interprets inline prompt examples too literally. Providing conversational matching triggers (e.g., “If prospect agrees, e.g., ‘Yes’”) causes the model to fixate on the exact phrasing rather than evaluating the semantic intent. This results in skipped dialogue steps.
  3. Prompt Inflation Sensitivity: Long, deterministic prompts containing multi-step dialogue rules are highly unstable in 2.1. The reasoning layer seems to prioritize overthinking structural transitions rather than strictly outputting direct responses.

Our Benchmarks:

  • gpt-realtime-mini (Legacy Snapshot): Highly predictable, immediate execution, stable state tracking, clean UX.
  • gpt-realtime-2.1-mini: Heavy conversation latency, structural hallucinations, severe instruction narration. Temporary reversion back to realtime-mini immediately restored 100% stability.

Request to OpenAI:
For structured, deterministic conversational state machines, the reasoning layer injected into the 2 and 2.1 branches acts as a severe UX bottleneck. We request either an extension of the legacy realtime-mini deprecation timeline or a dedicated feature flag/toggle to completely bypass the internal commentary phase for strict execution agents.

:folded_hands: