Codex GPT-5.4 Retirement Migration Guide (2026): Move to GPT-5.6 Terra or Luna Before August 31

As of August 2, 2026, OpenAI's Codex rate card carries a warning developers should not ignore: GPT-5.4 and GPT-5.4 mini will retire in Codex on August 31, 2026 for users signed in with ChatGPT. OpenAI's recommended replacements are direct: use GPT-5.6 Terra instead of GPT-5.4, and use GPT-5.6 Luna instead of GPT-5.4 mini.

This is not a panic migration. API access to GPT-5.4 is not affected by the Codex retirement notice, and Codex with your own API key is also not affected. But if your team runs Codex through ChatGPT Plus, Pro, Business, Enterprise, Edu, Gov, Health, or Teachers accounts, you have less than a month to update defaults, budgets, docs, and internal agent recipes.

What Is Actually Retiring?

The wording matters. OpenAI did not say every GPT-5.4 endpoint vanishes. The Codex notice says GPT-5.4 and GPT-5.4 mini retire in Codex for users signed in with ChatGPT. If your developers launch Codex from a subscription account and pick GPT-5.4 today, that model choice needs to move before August 31.

If you call the OpenAI API directly, your immediate migration pressure is lower. Still, I would not ignore the signal. Product defaults tend to move before platform defaults. When a model leaves Codex, tooling docs, examples, internal onboarding guides, and saved automations often lag behind. That is how teams end up debugging a "random" model error on Monday morning.

Pricing Table: Current Models Involved in the Migration

The interesting bit is that Terra is cheaper than GPT-5.4, while Luna is much cheaper than GPT-5.4 mini and also has a larger context window. That does not mean Luna should replace every mini workflow blindly. It means your old assumptions deserve a fresh eval.

Model Comparison: Terra vs Luna vs the Retiring Codex Defaults

A Practical Migration Plan

Start by splitting your Codex usage into three buckets. Do not make one global replacement and call it done.

1. Complex planning and risky edits: move from GPT-5.4 to GPT-5.6 Terra.
2. Routine implementation: test GPT-5.6 Luna first, then escalate failures to Terra.
3. Large-context repo work: compare Terra and Luna because both now support a 1,050,000-token context window.

The best setup is usually a two-model route. Terra decides when the task is underspecified or dangerous. Luna handles well-scoped work where the acceptance criteria are clear. This is boring engineering, but it saves money without making your agent dumb.

API Example: Route Codex-Style Tasks by Risk

Here is a small OpenAI-compatible pattern you can adapt for your own task runner. It defaults to Luna for routine work and bumps to Terra when the task looks risky.

curl https://api.openai.com/v1/responses \
  -H "Authorization: Bearer $OPENAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-5.6-terra",
    "input": "Plan a safe migration from GPT-5.4 Codex defaults to GPT-5.6 Terra and Luna. Return test steps and rollback criteria."
  }'

If you use KissAPI as a unified endpoint, keep the same routing idea and swap only the base URL and key used by your gateway. That gives you one place to test model behavior, track costs, and keep a fallback path if your primary provider quota gets tight.

Python Router Example

from openai import OpenAI

client = OpenAI(api_key="YOUR_API_KEY")

RISK_WORDS = ("auth", "payment", "database", "migration", "delete", "security")

def choose_model(task: str) -> str:
    lowered = task.lower()
    if any(word in lowered for word in RISK_WORDS):
        return "gpt-5.6-terra"
    return "gpt-5.6-luna"

def run_codex_task(task: str):
    model = choose_model(task)
    return client.responses.create(
        model=model,
        input=f"You are a careful coding agent. Task:\n{task}"
    )

print(run_codex_task("write unit tests for the invoice formatter"))

Node.js Router Example

import OpenAI from "openai";

const client = new OpenAI({ apiKey: process.env.OPENAI_API_KEY });

function chooseModel(task) {
  const risky = /auth|payment|database|migration|delete|security/i.test(task);
  return risky ? "gpt-5.6-terra" : "gpt-5.6-luna";
}

export async function runTask(task) {
  const model = chooseModel(task);
  return await client.responses.create({
    model,
    input: `You are a careful coding agent. Task:\n${task}`
  });
}

Migration Checklist Before August 31

• Find hardcoded model names. Search for gpt-5.4 and gpt-5.4-mini in shell aliases, editor settings, CI jobs, and internal docs.
• Update Codex defaults. Use Terra where GPT-5.4 was chosen for judgment, planning, or risky code changes.
• Benchmark Luna separately. It is cheap enough to be tempting, but test it against real tasks before making it the default.
• Watch output tokens. Agentic coding can spend more on output than teams expect, especially when it explains every step.
• Set budget alerts. OpenAI says Codex, ChatGPT Work, ChatGPT for Excel, and Workspace Agents can draw from the same agentic usage and credit pool when available on a plan.

One small warning: if you move from GPT-5.4 mini to Luna, do not judge the migration only by per-token price. Measure task completion rate, retry count, and review burden. A cheap model that needs three attempts is not cheap. A routing setup through KissAPI or your own gateway can help keep that data visible across models.

FAQ

Does GPT-5.4 disappear from the OpenAI API on August 31, 2026?

No. OpenAI's Codex notice says OpenAI API access and Codex use with your own API key are not affected by the August 31, 2026 Codex retirement notice.

Should GPT-5.6 Terra replace GPT-5.4 in Codex?

Yes, for most GPT-5.4 Codex workflows. OpenAI explicitly recommends GPT-5.6 Terra instead of GPT-5.4 for Codex users signed in with ChatGPT.

Should GPT-5.6 Luna replace GPT-5.4 mini in Codex?

Usually, but test it first. OpenAI recommends GPT-5.6 Luna instead of GPT-5.4 mini, and Luna costs $0.20 input and $1.20 output per 1M tokens with a 1,050,000-token context window.