How to Use OpenAI Models in OpenClaw in June 2026: GPT-5.5, Codex OAuth, and Safe Routing

OpenClaw’s June 2026 OpenAI story is cleaner than it looked in spring, but the naming still trips people up. As of June 9, 2026, npm shows openclaw stable at 2026.6.1 and beta at 2026.6.5-beta.6. The official docs now make the routing rule explicit: for OpenAI-backed agent turns, openai/* is the canonical model route, and the runtime and auth choice determine whether you are using native Codex subscription execution or direct API billing.

If you are still mixing up provider, model, runtime, and channel, stop and clean that up first. That confusion is the fastest way to get the wrong invoice, the wrong latency profile, or a broken voice setup.

1. Start from the current stable baseline

For production, the safe June 2026 default is still stable 2026.6.1. npm metadata on June 9 shows latest pointing to 2026.6.1 while beta points to 2026.6.5-beta.6. That matters because most OpenAI routing guidance in the docs assumes the newer openai/* model path and the bundled Codex runtime behavior that current releases repair automatically.

If you have not validated beta in your own workspace, stay on stable and use a separate test agent before you promote anything. For the broader release cadence and channel logic, see our OpenClaw update channels guide.

2. What openai/* means in OpenClaw now

The current OpenClaw provider docs say openai/* is the canonical OpenAI route. For agent turns, that route uses the native Codex app-server runtime by default. In other words, picking openai/gpt-5.5 is no longer just a model choice. It is also the normal way to say, “run this OpenAI agent turn through Codex,” unless you explicitly override runtime behavior.

The companion runtime docs are even more direct: the common ChatGPT/Codex subscription setup keeps the model reference as openai/*, stores auth as an OpenAI OAuth profile, and lets the bundled codex plugin run the embedded turn. That means provider, model, runtime, and channel are separate layers. If you treat them as one setting, you will eventually misconfigure production.

3. Use Codex OAuth when you want subscription-backed agent turns

If your goal is the normal ChatGPT or Codex-backed OpenClaw experience, the docs now recommend a simple path:

openclaw onboard --auth-choice openai
openclaw config set agents.defaults.model.primary openai/gpt-5.5
openclaw models list --provider openai

For headless servers or callback-hostile environments, OpenClaw documents the device-code variant:

openclaw models auth login --provider openai --device-code

OpenAI’s own help center currently says Codex works by signing in with your ChatGPT account and that availability depends on your ChatGPT plan and usage limits. That matches the OpenClaw docs: keep the public model ref on openai/gpt-5.5, let Codex OAuth handle auth, and verify the live runtime with /codex status or /codex models once the gateway is up.

4. Use API keys when you need direct billing or non-agent OpenAI surfaces

The same OpenClaw docs draw a clean line between agent turns and direct OpenAI Platform usage. Images, embeddings, speech, realtime voice, and other non-agent API surfaces still use funded OpenAI Platform access. The documented onboarding path is:

openclaw onboard --auth-choice openai-api-key
openclaw models list --provider openai

OpenClaw also supports a more explicit route if you intentionally want agent turns billed through the OpenAI API rather than your ChatGPT/Codex subscription: keep the model as openai/gpt-5.5 and select the provider/model runtime openclaw, or place a Codex-compatible API-key profile in your OpenAI auth order.

The docs also call out one trap that operators keep missing: openai/chat-latest is a moving alias for experiments, not the stable default. If you want predictable behavior, keep openai/gpt-5.5 as the production baseline.

5. The three routing mistakes that still break real deployments

First: assuming an openai/* model ref automatically means API billing. On current OpenClaw, it usually means the opposite for agent turns: Codex runtime by default.

Second: assuming Codex-backed chat auth also funds every OpenAI surface. The provider docs explicitly warn that Realtime voice still bills against OpenAI Platform credits. A team can have healthy Codex OAuth for chat and still fail voice setup until a funded API-key path exists.

Third: carrying forward legacy Codex refs. OpenClaw now documents openclaw doctor --fix as the repair path for older legacy Codex model refs, codex-cli/* refs, stale auth ordering, and old runtime session pins. If you upgraded from older spring builds, run the repair before you blame the model.

6. A practical June 2026 decision map

Goal Recommended route Why
Fastest default OpenAI agent setup openai/gpt-5.5 + Codex OAuth Matches current OpenClaw defaults and native Codex runtime behavior.
Predictable OpenAI API billing for agent turns openai/gpt-5.5 + explicit openclaw runtime or ordered API-key auth Keeps billing and auth behavior intentional instead of inherited.
Experimental ChatGPT-style instant alias openai/chat-latest Useful for experiments, but the alias moves and is not the production default.
Voice or realtime workflows Separate funded OpenAI Platform credits Codex subscription auth does not replace Platform billing for Realtime voice.

If you are comparing providers rather than standardizing on OpenAI, pair this with our Anthropic billing guide so your agent routing, auth, and spend controls stay coherent across both stacks.

Sources and verification

Need a production rollout instead of another weekend of routing fixes? ALL CLEAR DIGITAL helps teams choose the right provider mix, repair legacy OpenClaw model paths, and turn one-off agent setups into repeatable operating playbooks with clear cost and safety controls.