How to Choose an AI Model for Support Agents: Macha's GPT-5 Lineup

We've reworked the OpenAI lineup in Macha around exactly that trade-off. There are now three tiers — GPT-5.4 Mini, GPT-5, and GPT-5.4 — priced so that the right call is usually obvious: run the cheapest model that clears the bar, and reach for a heavier one only where the work genuinely needs it. This post is a practical guide to choosing a model per support agent: what changed, where OpenAI fits in Macha's wider model landscape, what each tier is actually good at, a worked credit-cost example, and how to assign models so you're not paying frontier prices to answer "where's my order?"

Quick reminder on credits. In Macha, a credit is the unit of an AI action, not a per-resolution fee. A credit cost is charged once per complete assistant response, and it varies by the model the agent is running. Across every model and provider in the platform the range runs roughly 0.5–9 credits per action; the three OpenAI models below sit at the affordable end. See the pricing page for current plans and allowances.

What changed

The lineup landed in two moves. First, we added GPT-5.4 as the top-capability OpenAI model and made GPT-5 the new general-purpose model, both with a large (1M-token) context window — and retired three older minis (GPT-5 Mini, GPT-4o Mini, and o4-mini), automatically migrating every agent that was on them to GPT-5.4 Mini so nothing broke. Then we cut the credit cost of the two larger models so you get materially more work out of the same allowance.

Here's where the three OpenAI tiers stand today:

The price cut matters more than it looks. Because GPT-5.4 dropped from 5 credits to 3, a Professional plan's 10,000 credits now stretches to ~67% more GPT-5.4 replies than before; GPT-5 dropping from 3 to 2 buys you ~50% more GPT-5 replies. Same plan, same bill — more headroom to put a capable model on the agents that earn it.

Where OpenAI fits in Macha's model landscape

Macha is multi-provider, so "which model should this agent run on?" is a slightly bigger question than the three OpenAI tiers. Alongside the GPT-5 family, Macha also offers Anthropic's Claude (Sonnet 4.5 and Sonnet 4) and Groq models — the Agents table even shows a provider icon next to each agent's model so a mixed fleet is easy to read at a glance. If you've followed the wider 2026 model conversation, you'll recognise the usual shortlist for support: the GPT-5 family, Claude Sonnet, and a fast/cheap option like Gemini Flash or a Groq-hosted model. Independent roundups such as eesel's "which LLM is best for customer support" land on the same split — a strong general model for replies, a long-context model for document-heavy work, and a cheap model for volume.

The OpenAI tiers are the focus here for two practical reasons. First, GPT-5.4 Mini is the platform default — it's where new agents start and where the retired minis were migrated, so it's the model most teams are already running. Second, the three OpenAI tiers map cleanly onto the cheap → balanced → frontier ladder that the rest of this guide is built around, which makes them the easiest place to learn the per-agent trade-off before you start mixing in Claude for, say, long-document policy work. (Picking between providers — Claude vs GPT for tone and long context — is a deeper topic; here we stay inside the OpenAI ladder, where the credit math is most directly comparable.)

How to think about picking a model

The instinct to default everything to the smartest model is understandable and wrong. Independent support-LLM comparisons keep landing on the same architecture: route simple work to small models and reserve frontier models for the genuinely hard cases. Anthropic's own ticket-routing guide is blunt about it — classification, extraction, and summarization are "well within the capability of smaller, cheaper models." And cost analyses like CloudZero's 2026 LLM pricing comparison note that model selection alone can swing your monthly spend by 10x. The skill isn't picking the best model; it's picking the cheapest model that still passes.

Macha makes that a per-agent decision instead of an account-wide compromise. A practical way to assign the tiers:

Start everything on GPT-5.4 Mini (1 credit)

This is the default for a reason. For the bulk of a real support queue — order-status lookups, password resets, "how do I…" FAQ answers, tagging and routing, ticket summaries — a fast mini model with a 400K context window is not a downgrade, it's the correct tool. At 1 credit per response it's also the only tier cheap enough to run across thousands of tickets without flinching. If an agent's job is triage and deflect, leave it here.

Promote to GPT-5 (2 credits) when replies get nuanced

When an agent is composing customer-facing replies that need tone, judgment, and several tool calls chained together — pull the order from Shopify, check the refund window, draft the response — GPT-5 is the balanced pick. It's the workhorse: smart enough for the messy middle of your queue, still cheap enough to run at volume. Most teams end up with their primary front-line resolution agent here.

Reserve GPT-5.4 (3 credits) for the hard 10%

GPT-5.4's edge is instruction-following under pressure: agents with strict "do this exactly once" rules, long runbooks, or complex branching logic where a smaller model occasionally colours outside the lines. Think a billing agent that must follow a precise refund-approval policy, or a multi-step escalation flow with conditional handoffs. You don't want this on your whole queue — you want it on the agents where a wrong step is costly.

Because the selector lives in each agent's config, your account can look like a sensible tiered system: a fleet of Mini-powered triage and FAQ agents feeding a couple of GPT-5 resolvers, with one or two GPT-5.4 specialists on the gnarly stuff. The Agents table shows a Model column so you can scan that distribution across your whole set at a glance — a fast way to catch the agent someone quietly bumped to GPT-5.4 "just to be safe."

A worked credit-cost example: tiered vs all-frontier

The "pick the cheapest model that passes" argument is easy to wave at and easy to ignore, so here's the actual arithmetic. Take a team handling 4,000 AI responses a month with a fairly normal mix:

• 70% triage — order status, FAQ deflection, tagging, routing, summaries → 2,800 responses
• 25% resolution — nuanced customer-facing replies with a few tool calls → 1,000 responses
• 5% complex — strict-policy billing, long runbooks, conditional escalation → 200 responses

Now price that two ways. "Play it safe" means running everything on the top model; tiered routing means matching each band to the cheapest model that clears its bar.

Same 4,000 responses, same hard tickets handled by the same frontier model — but tiered routing uses 55% fewer credits. In plan terms that's the difference between comfortably fitting inside a Professional 10,000 credits allowance and blowing past it (and reaching for top-up packs) two-thirds of the way through the month. The waste in the "safe" column is almost entirely the 2,800 triage tickets paying 3 credits for an answer a 1-credit model returns identically.

This isn't a Macha-specific quirk. Independent analyses of LLM model routing report bill reductions in the 40–85% range from sending routine work to small models and reserving frontier models for hard reasoning, with a 70/30 cheap-to-frontier split cutting the bill by roughly two-thirds — see digitalapplied's 2026 routing guide. The only difference in Macha is that you don't need a separate routing layer to capture it: the lever is the per-agent model selector you already have.

Where context window actually matters

Both GPT-5 and GPT-5.4 carry a 1M-token context window, and GPT-5.4 Mini carries 400K. That's not a vanity spec for support work. Context budget is what lets an agent hold a long Zendesk comment thread, the relevant knowledge-base articles, and the outputs of several tool calls in its head at once without losing the thread. If your agents work tickets with long histories, attached docs, or chunky knowledge sources, the larger window on GPT-5 / GPT-5.4 is a concrete reason to size up — independent of raw reasoning. For short, single-turn deflections, you'll never touch the ceiling and Mini is plenty.

Per-conversation overrides in chat

The per-agent model is the durable setting, but in the dashboard chat you can switch models for a single conversation from the top-right dropdown — handy for spot-checking how a heavier model handles a specific thread before you commit an agent to it.

Macha shows a toast when you do this, clarifying that the change applies only to the current conversation and pointing you to Settings if you actually meant to change the org default. It's a small guardrail against the classic "I changed it in chat and wondered why nothing else updated" trap.

Watch-outs and when not to size up

A few honest caveats so this doesn't read as "always buy the bigger model":

• More expensive ≠ better answers on easy work. On routing and FAQ deflection, GPT-5.4 will very often produce the same answer as Mini for 3x the credits. Paying up there is pure waste — the comparisons above all make this point. Size up only when you can name the failure a smaller model is making.
• Test before you standardize. Use a per-conversation override (or a Study) to see whether the cheaper tier actually fails on your traffic before assuming it will. Often it doesn't.
• Credits are per action, not per resolution. A chatty agent that takes six tool-calling turns to resolve a ticket costs six responses' worth of credits regardless of tier. Model choice is one lever; agent design (fewer, tighter turns) is the other.
• Retired models migrated automatically. If you were on GPT-5 Mini, GPT-4o Mini, or o4-mini, those agents are already on GPT-5.4 Mini — usually a straight upgrade, but it's worth a glance at any agent that depended on a specific older model's quirks.
• Enterprise bypasses credit checks, so the per-tier credit math is mostly a Starter/Professional planning concern; on Enterprise the decision is purely quality-and-latency.

Bottom line

The new lineup isn't about a single "best" model — it's about making the tiered approach that every serious support-AI guide recommends the default rather than something you have to engineer. Put GPT-5.4 Mini on the volume, GPT-5 on the resolutions, GPT-5.4 on the hard edges, and let the per-agent selector keep each piece of your queue on the cheapest model that still does the job. Macha sits on top of Zendesk, Freshdesk, Gorgias, and Front as an agent layer, so this model choice rides along with whatever helpdesk you already run.

Want to try it? Start a 7-day free trial, no credit card required, build an agent, and switch its model from the config card — or read the model docs for the full per-agent and per-conversation behaviour.

FAQ

What's the cheapest model in Macha, and is it good enough? GPT-5.4 Mini at 1 credit per response is the cheapest OpenAI tier and the default. For triage, routing, classification, summaries, and FAQ deflection it's not a compromise — those tasks sit well within a small model's range. Reserve pricier tiers for nuanced replies and strict multi-step logic.

What's the difference between GPT-5 and GPT-5.4? GPT-5 (2 credits) is the balanced general-purpose model — capable, runs well at volume. GPT-5.4 (3 credits) is the most capable, with notably stronger instruction-following for agents with strict rules or complex branching. Both have a 1M-token context window.

How many credits does each model cost? As of June 2026: GPT-5.4 Mini = 1 credit, GPT-5 = 2 credits, GPT-5.4 = 3 credits, charged once per complete response. Across all models and providers the range is roughly 0.5–9 credits per action. See pricing for current allowances.

Can different agents use different models? Yes — model is a per-agent setting in the agent's configuration card, and the Agents table shows a Model column so you can see the distribution. You can also override the model for a single chat conversation from the top-right dropdown.

What happened to GPT-5 Mini, GPT-4o Mini, and o4-mini? They've been retired. Every agent that was on them was automatically migrated to GPT-5.4 Mini, so no configuration broke.

Written by Abbas (Customer Support & AI, Macha) · Reviewed by Ankeet Guha (Co-founder & CTO) · Published 2026-06-24 · Last updated 2026-06-24.