A lot of the new "GPT 5.4 sucks in OpenClaw" posts are really config issues

A lot of the new "GPT 5.4 sucks in OpenClaw" posts are really config issues

Since the Anthropic change, a lot more people here are trying GPT 5.4 in OpenClaw.

Which makes sense. Claude Pro / Max inside OpenClaw was one of the biggest cost-saving setups, so now a lot of people are testing GPT 5.4 harder than before.

What I keep seeing though is people blaming the model for stuff that is mostly setup.

A lot of them are testing on old OpenClaw versions. Or with reasoning off. Or with weak thinking settings. Or without the right OpenAI path.

At that point, you’re not really testing GPT 5.4 properly.

You’re testing a crippled setup.

That’s also where the weird "it looks like it’s working but it’s not actually doing anything" feeling comes from.

If reasoning isn’t active, the bot is way more limited than people think. It can answer the last message. It can sound plausible. But as soon as the task needs actual multi-step reasoning, or something changes mid-flow, it falls apart fast.

That’s why some demos feel fake.
Not because GPT 5.4 is automatically bad in OpenClaw, but because the setup is bad enough that the bot never had much room to work with.

The biggest fixes for me were:

• update OpenClaw to at least 2026.4.5
• turn reasoning on
• keep thinking at least at medium
• use openai-responses
• enable block streaming if the bot is living in Telegram
• keep enough recent context so the conversation doesn’t degrade

Once those were fixed, GPT 5.4 felt way better.

Less fake progress.
Less random stopping.
Better continuity.

One important thing though: even with the right setup, GPT 5.4 still doesn’t feel exactly like Opus 4.6.

Opus 4.6 would sometimes take a lot more initiative. Sometimes that felt great. Sometimes it was honestly too much, and that freedom could also lead to mistakes.

GPT 5.4 feels a bit different. In my experience, it benefits more from validation and tighter steering on some steps.

Personally, I prefer that.

I’d rather have a model that needs a bit more checking, but stays more controllable, than one that takes too much initiative and occasionally goes off in the wrong direction.