OpenAI + OpenClaw: the GPT 5.4 setup that works

OpenAI and OpenClaw work well together, but GPT 5.4 only feels solid in production when a few specific settings are enabled. If you are trying to run a Telegram assistant, Discord helper, or internal business agent, the difference between a rough setup and a polished one usually comes down to protocol choice, reasoning, and streaming behavior.

This article is for solopreneurs and non-technical operators who want the OpenAI + OpenClaw stack without spending a night debugging config files.

Key takeaways

• In ClawRapid, GPT 5.4 uses openai-responses, not openai-chat.
• We set thinkingDefault to medium and reasoningDefault to on.
• We enable block streaming with text_end, 800 to 1200 char chunks, and 1000 ms coalescing.
• We keep 60000 recent tokens for better continuity.
• For Telegram, streaming configuration has an outsized effect on perceived quality.

Why OpenAI + OpenClaw can feel great or mediocre

The base stack is strong. OpenClaw supports direct OpenAI routes, and the OpenAI provider docs explicitly list openai/gpt-5.4 as a direct OpenAI Platform route (OpenClaw OpenAI docs). But the docs alone do not tell most people which settings matter most once the agent is live in a chat app.

That is where things usually go wrong.

A weak setup often has three problems at once:

1. the wrong API path
2. too little thinking
3. rough message delivery in chat

None of those make GPT 5.4 look broken. They just make it feel worse than it should.

The exact GPT 5.4 settings we use in ClawRapid

In lib/cloud-init.ts, ClawRapid configures GPT 5.4 with these values:

That setup is not theoretical. It is the version we already use because it behaves better in real chat.

Why openai-responses matters more than people think

The OpenClaw OpenAI provider docs state that openai/gpt-5.4 runs through the OpenAI Platform route, and OpenClaw forwards openai/* through the Responses path. That matters because the Responses API is the route OpenClaw is built around for current OpenAI models.

If you use an older or mismatched path, you are adding friction before you even get to prompt quality.

So if someone asks me what the minimum sane OpenAI OpenClaw setup is, openai-responses is the first line item.

Why thinking: medium and reasoning: on should be your floor

OpenClaw's thinking docs support levels from off up to xhigh, with medium sitting in the practical middle for reasoning-capable models (Thinking Levels docs). In ClawRapid, we set GPT 5.4 to:

• thinkingDefault: "medium"
• reasoningDefault: "on"

This is the floor, not an aggressive tuning choice.

When GPT 5.4 is pushed too low on thinking, answers can come back fast but thin. When reasoning is disabled, multi-step tasks and careful explanations are more likely to feel rushed. For a business assistant, that tradeoff is usually wrong.

If you are using OpenClaw for customer support triage, lead qualification, or a founder copilot, faster but sloppier is not a win.

Why block streaming changes the Telegram experience

OpenClaw's streaming docs make an important distinction: block streaming and preview streaming are separate layers. The same docs also say agents.defaults.blockStreamingDefault is off by default (Streaming + chunking docs).

That is one of the easiest ways to get an underwhelming setup.

In ClawRapid, we explicitly enable:

• blockStreamingDefault: "on"
• blockStreamingBreak: "text_end"
• blockStreamingChunk: { minChars: 800, maxChars: 1200 }
• blockStreamingCoalesce: { idleMs: 1000 }

This matters because Telegram punishes noisy output. Tiny fragmented messages feel messy, even if the model itself is fine. The OpenClaw streaming docs also note that Telegram preview streaming is skipped when Telegram block streaming is explicitly enabled, which helps avoid double-streaming behavior.

For most non-technical users, this is the hidden reason a DIY setup feels less polished than a managed one.

The practical decision framework

If you are setting up OpenAI + OpenClaw yourself, use this checklist:

Use this setup if you want the fastest path to something solid

• Use openai/gpt-5.4
• Keep the OpenAI route on the Responses path
• Set thinking to medium
• Keep reasoning on
• Turn block streaming on explicitly
• Use text_end for cleaner block boundaries
• Keep a large recent context window, like 60000 tokens

Revisit your config if any of these happen

• Telegram replies feel choppy
• Long conversations lose continuity
• The agent sounds shallow on multi-step tasks
• The model feels weaker in OpenClaw than in direct use elsewhere

How this compares with doing it manually

You can absolutely self-host OpenClaw and wire all of this yourself. If you are technical, that can make sense.

But for the ClawRapid audience, the real goal is not to understand every OpenClaw detail. The real goal is to have an assistant that replies cleanly, keeps context, and does not feel messy inside Telegram.

ClawRapid exists for people who want:

• a Telegram or chat-based assistant in about 60 seconds
• OpenAI and OpenClaw already configured to work together cleanly
• human support when something needs adjustment
• no-code delivery instead of config archaeology

That is also why this article is different from a generic OpenClaw setup guide. The point is not to explain every moving part. The point is to isolate the few settings that change the experience the most.

Where to change these settings

If you just want the config block, this is the part most people are looking for.

In practice, the exact structure can vary a bit depending on how your OpenClaw config is organized, but these are the settings that made the biggest difference for GPT 5.4 in chat:

{
  agents: {
    defaults: {
      blockStreamingDefault: "on",
      blockStreamingBreak: "text_end",
      blockStreamingChunk: { minChars: 800, maxChars: 1200 },
      blockStreamingCoalesce: { idleMs: 1000 },
      models: {
        "openai/gpt-5.4": {
          thinkingDefault: "medium",
          reasoningDefault: "on"
        }
      }
    }
  },
  models: {
    providers: {
      openai: {
        api: "openai-responses"
      }
    }
  }
}

If you are using Telegram, block streaming matters a lot more than most people expect. If you skip it, the model can still be fine, but the chat experience often feels much rougher than it should.

Common mistakes in OpenAI OpenClaw setups

Here are the mistakes I would avoid first:

That table alone fixes most self-inflicted setup issues.

What actually matters in production

If you forget everything else, keep these points:

• GPT 5.4 needs to be on the right OpenAI path
• thinking should stay at medium minimum
• reasoning should stay enabled
• block streaming should be turned on explicitly
• recent context should be large enough that the conversation stays coherent

That is what separates an assistant that is technically connected from one that actually feels good to use.

FAQ

Does OpenAI work well with OpenClaw?
Yes. OpenClaw supports direct OpenAI routes, and GPT 5.4 works well when the setup is tuned correctly. The big wins are using the Responses path, keeping thinking at medium, leaving reasoning on, and enabling block streaming for cleaner chat delivery.

What matters most for Telegram?
Streaming behavior matters a lot. Telegram exposes rough output quickly, so block streaming has a big effect on perceived quality. In practice, blockStreamingDefault: "on", text_end, and chunking around 800 to 1200 characters make replies feel much cleaner.

Is thinking: medium really necessary for GPT 5.4?
For most business-assistant use cases, yes. Lower settings can feel faster, but often give weaker multi-step reasoning and more shallow answers. medium is a good floor if you care about reliability more than shaving off a little latency.

Should I use ClawRapid or self-host?
If you are technical and want full control, self-hosting OpenClaw is fine. If you are a solopreneur, coach, freelancer, or operator who wants the stack working without tuning every config detail, ClawRapid is the more practical choice.

What to remember

OpenAI and OpenClaw are a good pair.

The problem is not compatibility. It is configuration.

If GPT 5.4 underperforms in an OpenClaw setup, the first things to inspect are the API route, thinking level, reasoning visibility, and streaming behavior. Those four choices shape most of the real user experience.

That is exactly why ClawRapid ships them preconfigured.

Want the OpenAI + OpenClaw stack without tweaking config files by hand? Deploy ClawRapid and start with GPT 5.4 already tuned for real chat use on Telegram and other messaging channels.

If you are comparing options, this is the simplest lens: self-host if you want total control, pick ClawRapid if you want the working setup without the configuration tax.