Choosing the Right AI Model

Why Model Choice Is The Biggest Lever

Model choice is the single largest determinant of quality and speed on Macha. The same instructions, same tools, and same workflow will produce noticeably different results depending on which model is behind the agent. It does not change what you pay: Macha bills per ticket, and a ticket costs the same whichever model handles it. So the choice is about getting the right answer at the right speed. Pick too heavy a model and a simple agent answers slower than it needs to. Pick too light a model and you ship an agent that misses obvious things and slowly erodes your team's trust.

The good news: model choice is not a gut call. There are three questions that get you to the right answer almost every time, and one cheap test that confirms it.

The Models

OpenAI GPT-5.6 Terra

The most capable model on Macha. Best for agents that need to reason across many pieces of context — long ticket threads, multiple tool results, or complex routing logic where the right answer depends on subtle signals. Use this for anything customer-facing that goes out without a human review step, anything billing-related, and anything where a wrong answer is expensive.

OpenAI GPT-5.6 Luna

A lighter GPT-5.6 model. A good fit when you want strong quality with quicker responses than Terra.

OpenAI GPT-5.4 (default)

The sweet spot for most agents, and the model new agents start on. Powerful enough to handle multi-step workflows with tool use and careful enough to follow detailed instructions. Start here unless you have a specific reason not to.

Stays on this tier: triage agents, routing agents, internal-only research agents, and most ticket-reply agents. Also a reasonable default for sub-agents called from a parent. Supports image vision.

OpenAI GPT-5

A solid general-purpose model, still available for agents that were built and tuned on it. Supports image vision.

OpenAI GPT-5.4 Mini

Fast. Excellent for high-volume workflows where the per-step reasoning is straightforward — categorising tickets by tag, looking up an order ID, summarising a thread, transcribing audio. Strong instruction-following for a mini model, and it supports image vision.

One caveat to know: mini models occasionally drop complex nested objects when calling tools with deeply nested parameter schemas. If you build a custom tool that takes a big nested config object and notice the mini model is calling it with fields missing, that is the cause — switch to GPT-5.4 for that specific agent or restructure the tool to take flat parameters.

Anthropic Claude Sonnet 5 and Claude Sonnet 4.5

Strong at nuanced writing, careful reasoning, and following long, complex instructions without skipping rules. Often the best choice for customer-facing replies where tone and empathy matter, and for agents that work from long, structured policy documents. Supports image vision.

If you have an instruction set with twelve numbered steps and several conditional branches, Claude tends to follow the structure more reliably than equivalent OpenAI models. The trade-off is slightly less consistent tool use in some edge cases.

GPT OSS 120B and GPT OSS 20B

OpenAI's open-weight models. Low latency — useful when you need a snappy response, like a Slack bot that has to feel quick in a real-time conversation. Trade-off: they are text-only and do not read images, so use them for text-only, latency-sensitive work.

The Three-Question Decision Framework

Walk through these in order. The first one that returns a hard answer settles the choice.

1. Does this agent send anything to a customer without human review? If yes, default to GPT-5.6 Terra, GPT-5.4 or Claude Sonnet. One bad reply at scale costs far more trust than any speed gain from a lighter model. The internal-notes-only testing pattern (covered in Testing Best Practices) lets you defer this decision until you have proof the agent's drafts are good enough.
2. Does this agent need to read images? If yes, exclude the GPT OSS models. The other OpenAI models and Claude work.
3. Is the workflow simple and high-volume? If yes — pure read tools, simple categorisation, internal-only writes — GPT-5.4 Mini is usually enough and answers faster. The classic example is a triage agent that runs on every new ticket.

If none of those apply: stay on GPT-5.4. It is the default for a reason.

Validate Your Choice With A Side-By-Side Test

The cheapest experiment in agent design is running the same agent on two models in parallel for a small batch of test cases. Do this before committing to a model in production.

Concretely: create the agent on the model you think is right. Use the Test Run feature to run it against ten varied real-world inputs (mix of easy, hard, and edge cases). Note the outputs. Now duplicate the agent, switch only the model, and run the same ten inputs. Compare side by side.

If the lighter model's outputs are indistinguishable from the heavier one's, you get faster responses at no quality cost. If the lighter one is meaningfully worse, you have proof that the heavier model is worth it. Either outcome is a win — you stop making the choice on intuition.

Changing Models On A Live Agent

You can change an agent's model at any time. The change applies to the next conversation — in-flight conversations finish on whatever model they started on. There is no migration step, no compatibility check, no downtime. This means it is genuinely safe to A/B model choice on a live agent: pick a small subset (one brand, one tag), switch the model, and watch.

If quality drops, switch back. If quality holds and responses get faster, keep the new choice and roll out to wider traffic.