OpenAI is ending support for GPT‑4o, GPT‑4o mini, GPT‑4.1 and GPT‑4.1 mini on January 14 2027, with the deprecation notice sent on October 15 2026. Practitioners who have built services, agents, or pipelines around these models now have just three months to migrate to the recommended replacement, GPT‑6 Luna, which introduces different reasoning and tool‑call semantics.
What Changed
The retirement applies to the base models, their dated snapshots, and every fine‑tuned variant built on top of them. OpenAI’s policy normally promises at least six months’ notice for generally available models, but an exception was made here, leaving developers with half the expected lead time. The List models API now returns a shutdown_date field, making future retirements discoverable programmatically.
Architectural and Implementation Implications
GPT‑6 Luna supports the same token‑window size (up to 1.05 million tokens) and structured outputs, but its API behavior differs in several key ways:
- Reasoning effort defaults to
medium. To use function calling via Chat Completions, callers must setreasoning_effort=none. - Combining reasoning with tool calls requires the separate Responses API.
- Enabling reasoning disables parameters such as
temperature,top_p,top_logprobs, andlogprobs, which were available on GPT‑4.1.
These changes mean that a simple model‑ID swap is insufficient; code that relies on the disabled parameters or on the previous reasoning‑tool coupling must be updated. OpenAI suggests using its Docs skill in Codex to rewrite API calls, but that automation does not cover behavioral differences, so thorough testing is required.
Operational and Cost Considerations
Luna’s pricing is $0.10 per million input tokens and $0.50 per million output tokens, cheaper than the $0.20/$1.20 rates of GPT‑5.6 Luna. However, once a request exceeds 272 000 input tokens, the input rate doubles and the output rate increases by 50 % for the entire request. For Retrieval‑Augmented Generation (RAG) pipelines or large‑code‑base agents, this tiered pricing can erode the nominal cost advantage.
Fine‑tuning on Luna is not available. New self‑serve fine‑tuning jobs close on January 6 2027, eight days before the GPT‑4 models retire, and existing fine‑tuned models will continue to serve inference until the final shutdown date. Teams that rely on custom fine‑tuned variants must either rebuild them on a supported model before the cut‑off or negotiate dedicated capacity with OpenAI’s sales team.
Security and Reliability Outlook
The deprecation does not introduce a new vulnerability, but the shortened migration window raises reliability risk. Applications that fail to switch before the shutdown will experience API errors, potentially causing service outages. Monitoring the shutdown_date field and setting alerts for deprecation warnings can mitigate surprise failures. Additionally, the change in reasoning semantics may affect prompt injection surface area, so security reviews should include new model behavior.
Related CloudNinjas coverage: AI engineering.
What This Means For Practitioners
Teams should take immediate action:
- Audit all production workloads that reference the retiring GPT‑4 models.
- Update integration code to handle Luna’s reasoning and function‑calling constraints, and replace disabled parameters with compatible alternatives.
- Run performance and cost benchmarks for token‑heavy workloads, especially those crossing the 272 k input‑token threshold.
- Plan migration of any fine‑tuned models before the January 6 2027 cut‑off, or engage OpenAI sales for extended capacity.
- Incorporate the
shutdown_datefield into CI/CD checks to catch future retirements early.
By treating the retirement as a forced upgrade rather than a simple version bump, engineers can avoid unexpected downtime and keep cost projections accurate.
