開發者平台
OpenAI Retires 13 Legacy Model Snapshots at Once, with API Calls to Five Early Codex Models Now Failing Hard
OpenAI shut down 13 legacy model snapshots on July 23, including five GPT-5-series Codex models and older computer-use and deep-research endpoints. The official replacement models are not aliases with identical behavior or cost, so existing agents need to be reevaluated—not merely updated with a new model string.
Following its previously announced deprecation schedule, OpenAI discontinued 13 API model snapshots on July 23. The models with the greatest impact on coding agents are `gpt-5-codex`, `gpt-5.1-codex`, `gpt-5.1-codex-max`, `gpt-5.1-codex-mini`, and `gpt-5.2-codex`: requests still pinned to these IDs will fail rather than being automatically routed to newer models. This retirement wave also covers two `chat-latest` aliases, two deep-research snapshots from 2025, and endpoints including `computer-use-preview-2025-03-11`.
OpenAI is primarily directing general Codex and chat workloads to GPT-5.5, mini Codex and legacy computer-use workloads to GPT-5.4 mini, and the two deep-research models to GPT-5.5 pro. These mappings are not drop-in replacements: reasoning effort, tool-call decisions, output formats, context pricing, and token consumption per agent loop may all change. In particular, after research workflows move to the pro tier, teams need to recalculate the cost model for long-running search and synthesis tasks. Browser agents must also be retested for clicking, waiting, error recovery, and safety confirmations.
The shutdown was announced as early as April 22 and was not unexpected, but the community continued reminding developers ahead of the deadline to inspect CI configurations, environment variables, model routers, and third-party IDE plugins. Teams should also distinguish between a hard API shutdown and the softer guidance of a “deprecated” label in Codex product interfaces, as the two may involve different models and dates. The next priority is not merely confirming that an endpoint responds, but comparing task success rates, side effects from tool operations, latency, usage limits, and actual billing. If model IDs are scattered across multiple services, teams should establish a centralized version inventory and conduct regular deprecation drills.