Macha Credits Explained — How AI Agent Billing Works
Macha doesn't bill per seat or per ticket — it bills in credits, and the model is simpler than most usage-based pricing once you know the two rules. Understanding it takes about two minutes, and it changes how you build agents cost-effectively.
Watch the quick explainer
Rule 1: one credit charge per complete response
Every time an agent generates a complete response, it uses credits — once. Not per tool call, not per step in its instructions, not per message it reads. The whole back-and-forth an agent does to handle a request — read the ticket, search knowledge, call a tool, decide, reply — is billed as a single response.
That's a big deal, because it means a complex agent that uses five tools to resolve a ticket costs the same as a simple one that just replies: one response. You're paying for outcomes the agent produces, not for how hard it worked to get there.
Rule 2: the cost depends on the model
How many credits a response costs is set entirely by the model you've chosen for that agent. Lighter, faster models cost less; stronger models cost more. Here's the full table:
| Provider | Model | Credits / response |
|---|---|---|
| OpenAI | GPT-5.4 Mini (default) | 1 |
| OpenAI | GPT-5 | 3 |
| OpenAI | GPT-5.4 | 5 |
| Anthropic | Claude Sonnet 4.5 | 9 |
| Anthropic | Claude Sonnet 4 | 9 |
| Groq | Llama 3.3 70B | 1 |
| Groq | Llama 3.1 8B | 0.5 |
| Groq | Mixtral 8×7B | 0.5 |
So a response can cost anywhere from 0.5 to 9 credits depending on which model you pick — your call, based on the task. You set the model per agent on its configuration screen.
What this means in practice
For most Zendesk workflows — triaging tickets, drafting replies, summarizing content, updating fields — GPT-5.4 Mini at 1 credit per response is all you need. It's fast, affordable, and handles those tasks well, which makes it the right default for high-volume support operations.
You step up to a stronger model only when the work demands it. An agent with a long, detailed set of instructions and many steps to remember will do better on GPT-5 (3 credits) — strong quality at a reasonable cost — or one of the premium models for genuinely hard reasoning. The point is you match the model to the task instead of paying premium rates for simple work.
How credits come in your plan
Each plan includes a monthly pool of credits:
- Starter — $299/mo, 3,000 credits.
- Professional — $699/mo, 10,000 credits.
- Enterprise — unlimited.
Credits reset each billing cycle, and if you need more mid-cycle, one-time top-up packs are available (they don't expire). One more thing worth knowing: **indexing your knowledge — building embeddings for your sources — is not charged against credits.** You only spend credits when an agent actually responds.
Keeping costs predictable
Because cost is "responses × the model's rate," there are exactly two levers:
- Pick the right model per agent. This is the big one. A high-volume triage agent on GPT-5.4 Mini (1 credit) instead of a premium model (9 credits) is a 9× difference for work that doesn't need it.
- Watch your volume. More responses = more credits. Agent Analytics shows you which agents are generating the most responses and on which models, so you can right-size.
Unlike a per-resolution meter that climbs with every ticket, a credit plan is a fixed monthly pool — so you always know your ceiling.
A worked example: what a busy month costs
Put real numbers on it. Say a support team automates 2,000 ticket actions a month — triage, summaries, and replies — all on GPT-5.4 Mini at 1 credit each. That's 2,000 credits, comfortably inside Professional's 10,000 credits. Swap a few hundred of those to GPT-5 (3 credits) for one complex agent and you're still well under the pool.
Now compare that to a per-resolution tool charging ~$1 per outcome on the same 2,000 actions — that's $2,000, and it climbs every month volume rises. The credit model's advantage isn't just the unit cost; it's the predictability: you're drawing from a fixed monthly pool, the cheap default model stretches it a long way, and a heavy month doesn't blow up your bill. The two levers — model choice and volume — are both visible and in your control.
Frequently asked questions
How are Macha credits charged? Once per complete AI response — not per tool call or per step. The amount depends on the model the agent uses.
How much is a response? 0.5 to 9 credits depending on the model; the default (GPT-5.4 Mini) is 1.
Does using more tools cost more? No — a response that uses several tools still counts as one response.
Is indexing my knowledge base charged? No — embedding and indexing your sources isn't billed; only agent responses use credits.
What if I run out mid-month? Buy a top-up pack (they don't expire), or your pool resets at the next billing cycle.
Which model should I default to? GPT-5.4 Mini (1 credit) for most support workflows; step up to GPT-5 for complex, multi-step agents.
The bottom line
Macha billing is two rules: one charge per complete response, at a rate set by the model. Default your agents to GPT-5.4 Mini for everyday support work, reserve stronger models for genuinely complex agents, and your credits stretch a long way — on a predictable monthly pool rather than a meter that grows with every ticket.
See it for yourself: start free and watch how few credits real workflows use. 7-day free trial, no credit card required. Start free.
