OpenAI sets 2027 retirement dates for older API models
OpenAI gives developers two deadlines for migrating older coding and speech models, with different replacement paths for each workload.
OpenAI published two API deprecation notices on October 1 covering older coding, general purpose and speech models. The affected models are scheduled for removal in 2027, leaving developers time to plan migrations.
Two deadlines to track
GPT 5.3 Codex and GPT 5.1 are scheduled to leave the API on April 1, 2027, with GPT 6 Sol recommended as their replacement. GPT 5.4 Nano has the same deadline, with GPT 6 Luna recommended.
The speech deadline is January 6, 2027. It covers TTS 1, TTS 1 HD and the GPT 4o mini TTS snapshots dated March 20 and December 15, 2025. OpenAI recommends GPT Realtime 2.1 Mini for all four in its official notices.
Deprecation begins the transition period. OpenAI says access ends at shutdown, so the announced dates belong on migration schedules.
Speech migration deserves its own review
The replacement model documentation describes GPT Realtime 2.1 Mini as a reasoning model for voice interactions that accepts audio and text. It supports WebRTC, WebSocket and SIP connections.
OpenAI directs speech developers to its Realtime API guide. That guide describes sessions that maintain conversation state and handle audio turns, tools and interruptions. Its browser example uses WebRTC, while server applications can connect through WebSocket.
For teams maintaining a narration pipeline, this creates a practical review task. Confirm which replacement workflow fits the product before estimating the engineering work. A deadline and recommended model alone cannot establish compatibility with an existing application.
Start with the application contract
OpenAI’s evaluation guidance recommends tests that reflect actual tasks and production traffic, with human review to calibrate scoring. A migration review should compare the behaviors users depend on, rather than rely only on a model’s broad benchmark position.
A useful first pass is to inventory configured models, identify the owner of each affected integration and define acceptance criteria. Speech teams can include pronunciation, timing and output handling. Coding teams can examine tool behavior and response formatting.
These are planning considerations, not results from independent API testing. Teams should measure quality, latency and cost in their own workload before choosing a cutover date.
The same planning discipline applies to other providers. ByteForward’s Claude Sonnet 4.5 retirement coverage explains a separate Anthropic deadline.
Original AI generated conceptual illustration of a planned software migration. Created for ByteForward.
