openclaw - 💡(How to fix) Fix [Bug] GPT-5.4 nano no longer survives as nano in Codex runtime (2026.4.5) [1 comments, 1 participants]

Summary

On OpenClaw 2026.4.5 in our installation, openai-codex/gpt-5.4-nano no longer behaves as a stable/real nano route in the Codex runtime.

We see two distinct behaviors depending on the call path:

1. In the current runtime, requesting nano can resolve to provider=openai-codex but the runtime reports model=gpt-5.4 instead of nano.
2. In another production call path (YouTube analysis), Crack observed nano falling through to OpenRouter/Sonnet instead of staying on Codex.

The key point is that nano is not surviving as a real nano route in our installation.

What is verified in our installation

• OpenClaw version: 2026.4.5
• openclaw models list --agent crack --json shows:
  • openai-codex/gpt-5.4-mini => available/configured
  • openai-codex/gpt-5.4-nano => configured/missing
• Mini works in Codex runtime.

Repro 1: nano request in Codex runtime does not come back as nano

Command (with session override already set to nano on agent:crack:main):

openclaw agent --local --agent crack --session-id 07ed6845-164b-44ff-8cb9-ef3169580945 --json --message 'Responde exactamente OK_NANO y nada más.' --timeout 120

Observed output in our installation:

• provider=openai-codex
• model=gpt-5.4
• response text OK_NANO

So the request does not preserve nano as the model identity.

Repro 2: production call path observed by Crack

In the YouTube analysis runtime path, Crack observed nano still failing and falling through to OpenRouter/Sonnet instead of Codex real.

Expected

If nano is supported, the runtime should either:

• execute as nano and report nano honestly, or
• fail explicitly with a precise unsupported-model error.

It should not silently re-route through an unrelated model or provider.

Related history

• The prior feature work to add mini/nano support is tracked in #50265.
• This report is about the current runtime behavior/regression in 2026.4.5, not the original feature request.

Ask

Please confirm whether:

• nano is intentionally no longer a first-class runtime route,
• the Codex runtime is silently normalizing nano to gpt-5.4, or
• there is still a bug in model resolution/fallback that can cause nano to drop into OpenRouter/Sonnet.