What's the best AI model for an ad-ops agent?

There isn't one — it depends on the agent. Put a strong or premium model on the orchestrator (plan quality drives everything downstream), the cheapest fast model on safety (structured checks, no reasoning needed), a long-context model on research, and start the rest on a cheap or strong tier, moving individual agents to premium only where you see output you'd reject.

Should every agent use the same model?

No — that's the most common mistake. Running a frontier model everywhere overpays for routine steps; running a cheap model everywhere starves the planner of the judgment it needs. Per-agent model choice lets you match cost to what each agent actually does.

Is a more expensive model always better?

Not for every task. Premium models earn their price on judgment-heavy, multi-step work — planning, hard optimization calls, creative output. For structured checks, simple data pulls, or high-volume routine steps, a cheap model performs the same at a fraction of the cost.

How do I know if I should upgrade an agent's model?

Watch for a pattern, not a single bad output: if you're consistently rejecting a specific agent's proposals or rewriting its output, that's the signal to move it up a tier. If everything it produces is fine, the current tier is already right — don't upgrade pre-emptively.

Do new AI models automatically become available?

In Agent Planners, yes — the Agents page pulls the live model gateway catalog and surfaces new chat models as providers ship them, correctly priced from the real gateway rate, without waiting for a product update.

Playbooks

How to choose an AI model for each agent: a decision framework

TL;DR

The mistake most teams make is picking one model for every agent — either overpaying by running a frontier model on every routine step, or underpaying by starving the planner of the reasoning quality it actually needs. The right approach is per-agent: a strong model where judgment drives the outcome, a cheap one everywhere execution is mechanical, and reasoning for the analysis that has to actually be correct. This is the decision framework, with real current pricing.

Start from what each agent actually does

Agent	What it does	Where model quality matters
Orchestrator	Decomposes your goal into a plan and delegates steps	Highest — a weak plan means every downstream step executes the wrong thing correctly
Analytics	Reads and reasons over Google Ads, GA4, BigQuery data	High for judgment calls (what's the anomaly, what's noise); low for simple pulls
Optimization	Proposes campaign, keyword, budget and PMax changes	High — this is what a human reviews and approves; bad reasoning here wastes an approval cycle
Creative	Writes ad copy, asset descriptions, video scripts	Medium-high — quality is subjective but visible immediately
Safety	Risk checks and stale-state validation before writes	Low — structured yes/no checks at low temperature, latency-sensitive
Research	Web research, competitor teardowns, SERP analysis	Medium — benefits from a large context window more than raw reasoning power

This maps directly to Agent Planners' own default preset: quality-critical agents get a strong model, safety stays cheap, research gets a long-context model — see GPT-5.6 Luna is now the default for the current specifics.

The three-tier framework

Cheap / fast tier (e.g. GPT-5.6 Luna, Gemini 3.5 Flash, Claude Haiku 4.5, DeepSeek Flash) — for safety checks, simple data pulls, and any step where the answer is either obviously right or gets caught downstream anyway. Note that “cheap” and “weak” aren't the same thing: GPT-5.6 Luna is priced in this tier at $0.20/$1.20 per 1M tokens but is strong enough to be Agent Planners' own default planner.
Strong tier (e.g. Claude Sonnet 5, Gemini 2.5 Pro, DeepSeek V4 Pro) — the balanced choice for most quality-critical agents when you want headroom above the cheap tier without premium pricing.
Premium tier (e.g. Claude Opus 5, Claude Fable 5, GPT-5.5 Pro) — 25-150 credits per 1K input tokens. Reserve for the orchestrator on genuinely hard, multi-step goals, or for creative work where output quality is the entire point.

A concrete starting point

1Put the orchestrator on a strong or premium model — plan quality determines everything downstream, and this is the single highest-leverage model choice you'll make.
2Put safety on the cheapest fast model available — it's running structured, low-temperature checks where a frontier model buys you nothing.
3Put research on a model with a large context window rather than the highest raw reasoning score — synthesis across many sources benefits more from context length.
4Start analytics, optimization and creative on a cheap or strong model, and only move a specific one to premium if you're seeing plans you'd reject or copy you're rewriting.
5Re-check monthly: model pricing shifts, and the cheap tier's capability keeps rising — a model you moved to premium six months ago might do fine on strong today.

Don't guess the price — check it

Model pricing varies by more than 100× across the catalog, and the gap between two models with similar-sounding names can be large. Rather than estimating, check the current rate before committing an agent to a premium model.

The model pricing & credits ranking lists every selectable model's credit rate, cheapest to priciest, refreshed daily from the real gateway price.
The AI model pricing calculator is the same data in a filterable/sortable tool.
Watch for new arrivals — Agent Planners auto-surfaces new chat models as providers ship them, so a genuinely better cheap-tier option can appear without you having to go looking for it.

One goal, several models, no contradiction

A single task run can legitimately call five different models — a premium planner, a cheap analytics pull, a strong optimization proposal, a cheap safety check, a long-context research step — because each agent's model is chosen independently. That's the point of per-agent choice: you're not trading off quality against cost once for the whole system, you're making that trade six separate times, each at the setting that actually fits the job.

Frequently asked questions

What's the best AI model for an ad-ops agent?: There isn't one — it depends on the agent. Put a strong or premium model on the orchestrator (plan quality drives everything downstream), the cheapest fast model on safety (structured checks, no reasoning needed), a long-context model on research, and start the rest on a cheap or strong tier, moving individual agents to premium only where you see output you'd reject.
Should every agent use the same model?: No — that's the most common mistake. Running a frontier model everywhere overpays for routine steps; running a cheap model everywhere starves the planner of the judgment it needs. Per-agent model choice lets you match cost to what each agent actually does.
Is a more expensive model always better?: Not for every task. Premium models earn their price on judgment-heavy, multi-step work — planning, hard optimization calls, creative output. For structured checks, simple data pulls, or high-volume routine steps, a cheap model performs the same at a fraction of the cost.
How do I know if I should upgrade an agent's model?: Watch for a pattern, not a single bad output: if you're consistently rejecting a specific agent's proposals or rewriting its output, that's the signal to move it up a tier. If everything it produces is fine, the current tier is already right — don't upgrade pre-emptively.
Do new AI models automatically become available?: In Agent Planners, yes — the Agents page pulls the live model gateway catalog and surfaces new chat models as providers ship them, correctly priced from the real gateway rate, without waiting for a product update.