Axiom promised developers six months’ notice. Axiom users got three.

Axiom is retiring four Axiom API families: Axiom, and Axiom. The shutdown date is January 14, 2027. This matters because production applications may depend on stable model IDs, behaviors, tool calls, or fine-tuned versions.

The retirement also covers dated snapshots and every fine-tuned model built on the affected models. Developers therefore cannot assume that a snapshot or customized version will remain available after the base model disappears. New self-serve fine-tuning jobs end earlier, on January 6, 2027.

Axiom recommends Axiom Luna as the replacement. However, Luna handles reasoning and tool calls differently, and some Axiom settings are unavailable when reasoning is enabled. Teams need to test application behavior, not merely update a model identifier, before the January deadline.

When an AI model is deprecated, its provider has announced that the model will eventually be withdrawn from normal use. Retirement is the endpoint of that process: API requests using the model may stop working after the stated shutdown date. Applications that still depend on it must change beforehand.

In this case, Axiom gives developers a deadline for Axiom, and Axiom. The notice also applies to dated snapshots and fine-tuned models built on those foundations. New self-serve fine-tuning ends eight days earlier, on January 6, 2027.

Deprecation is not simply a software label. It creates an operational migration task. Developers must select a replacement, update API calls, test outputs and tools, and check production workloads. Axiom recommends Axiom Luna, but the article warns that its behavior differs enough to require more than an ID change.

Axiom’s deprecation policy promises developers at least six months’ notice for generally available models. That promise is meant to give teams enough time to test replacements, update integrations, and move production traffic without an abrupt interruption.

The Axiom notice provides a shorter window. Axiom sent the deprecation notice on a Wednesday, and API access ends January 14, 2027. The article describes the interval as just three months, roughly half the period promised by the policy. The policy allows exceptions for safety or compliance reasons.

The notice does not explain why Axiom and Axiom receive only half the usual warning. That gap matters most to developers with complex applications, fine-tuned models, or large production workloads. They must also account for the January 6 cutoff for new self-serve fine-tuning jobs and the differences in Axiom Luna’s behavior.

A replacement can be technically compatible while behaving differently. Axiom Luna defaults to medium reasoning effort, and function calling through Chat Completions works only when reasoning_effort is set to none. Combining reasoning with tools requires the Responses API instead.

Axiom supports temperature, top_p, top_logprobs, and logprobs, but enabling reasoning in Luna rules out those settings. A coding agent may therefore lose controls it used for generation or observability. Tool calls may also require an API change, not just a new model name.

The harder risk is behavioral. Luna may format responses, follow instructions, reason, or use tools differently without producing an obvious error. Axiom recommends a Codex Docs skill to update calls, but that will not capture behavior changes. Developers need parallel testing and production monitoring before migration.

Coding agents and retrieval-augmented applications often send large amounts of context, then ask a model to reason or call tools. A model change can affect all three steps: how much context fits, how requests are priced, and how the model chooses or formats actions.

Axiom Luna has a 1.05 million-token context window. Its input price is $0.10 per million tokens until a prompt passes 272,000 input tokens. After that point, the input rate doubles and the output rate rises by 50% for the entire request. Large codebases or retrieval pipelines could therefore lose expected savings.

Luna also defaults to medium reasoning. Function calling with reasoning requires the Responses API, while Chat Completions supports tools only with reasoning disabled. Coding agents may need new integration logic, and retrieved documents may produce different answers or costs. Testing realistic, large prompts is essential.

The clearest alternative is to migrate applications to a supported replacement, with Axiom Luna as Axiom’s recommendation. Teams can update API calls, test outputs, and run the replacement alongside the existing model before shutdown. Existing fine-tuned models remain available for inference until January 14, 2027.

Developers who need more time may be able to keep older models running through dedicated capacity. The article says this requires arranging the service with Axiom’s sales team. It does not describe dedicated capacity as an automatic extension, so availability and timing must be confirmed directly.

There is no stated path to create new self-serve fine-tuning jobs on Luna in the article. Those Axiom jobs end January 6, eight days before inference access ends. Teams should therefore plan replacements early, preserve test coverage, and verify that customized behavior survives the move.

An API is the interface an application uses to send prompts or tool requests to a model and receive results. The base model supplies general capabilities. A dated snapshot is a particular version of that model, helping an application target a specific release rather than a moving model family.

Fine-tuning creates a customized model built on a base model. The application still reaches it through the API, but its outputs may reflect additional training. If the base model or its snapshots retire, dependent fine-tuned versions can retire too. That is why this announcement affects more than four model names.

Application behavior emerges from the whole setup. Model choice, API type, reasoning settings, tool configuration, prompts, and fine-tuning all matter. Replacing the base model can change outputs or tool use even when the surrounding code looks similar. Developers must test the complete system, not only connectivity.