GPT-5.4 Mini Is Here: What It Means for AI-Powered Customer Support
OpenAI just released GPT-5.4 Mini — the latest in their mini model line. It's faster at reasoning, handles longer conversations, and costs only slightly more than GPT-5 Mini. For teams running AI agents on customer support, this is the most meaningful model upgrade since GPT-5 Mini launched. Here's what changed and whether you should switch.
What's new in GPT-5.4 Mini
GPT-5.4 Mini is OpenAI's newest model in the mini tier — designed for high-volume use cases where cost matters but you don't want to sacrifice quality. The headline improvements:
- 400K context window — 3x larger than GPT-5 Mini's 128K. Your agent can process very long ticket histories, multi-turn conversations, and large knowledge-base documents without running into context limits.
- Stronger reasoning — better at following complex, multi-step instructions. Classification decisions are more accurate. Multi-tool workflows (read ticket → check order → apply business rules → draft response) produce fewer errors.
- Image vision — like GPT-5 Mini, it can see and analyse image attachments. Screenshots, product photos, receipts — the agent describes what it sees and acts on it.
- Same API, same endpoint — drop-in replacement. No code changes needed. Just switch the model ID in your agent settings.
GPT-5.4 Mini vs GPT-5 Mini: should you switch?
| Feature | GPT-5 Mini | GPT-5.4 Mini |
|---|---|---|
| Context window | 128,000 tokens | 400,000 tokens |
| Credit cost | 1 credit | 1.25 credits |
| Image vision | Yes | Yes |
| Reasoning quality | Good | Better |
| Speed | Fast | Fast |
Switch if:
- Your agents handle complex, multi-step workflows (WISMO automation, escalation logic, multi-tool chains)
- Your tickets have long conversation histories (10+ back-and-forth exchanges before the agent runs)
- You need the agent to follow nuanced instructions accurately (e.g., different response templates based on customer tier, language, and product category)
- The 25% credit increase is within your budget
Stay on GPT-5 Mini if:
- Your agents do simple tasks — tagging, routing, short replies
- Ticket conversations are typically short (1-3 messages)
- You're optimising for cost per resolution
How to switch
In Macha, each agent has its own model setting. You can switch one agent at a time — no global change needed:
- Open the agent's settings page
- Click the model selector (top right)
- Select GPT-5.4 Mini
- Save
The change takes effect on the next message or trigger. Existing conversations continue with their original model — only new interactions use the updated one.
You can also mix models across agents: your high-volume triage agent stays on GPT-5 Mini (1 credit), your complex WISMO agent upgrades to GPT-5.4 Mini (1.25 credits), and your escalation agent uses Claude Sonnet 4.5 (9 credits) for the hardest cases.
What about Claude Sonnet?
Claude Sonnet 4.5 (9 credits) is still the strongest model available for complex reasoning and nuanced writing. GPT-5.4 Mini doesn't replace it — but it narrows the gap. For many workflows that previously needed Sonnet's reasoning power, GPT-5.4 Mini may now be sufficient at 1/7th the credit cost.
The right comparison isn't "which is better" — it's "which is good enough for this specific agent's job." Test both on a sample of your hardest tickets and compare the output quality. If GPT-5.4 Mini handles 90% of them well, save the Sonnet budget for the 10% that truly need it.
