Ask five marketers which AI model they use and you will get five answers, none of them backed by a reason. "It's the one my team already had." "It's the one that went viral." That is not a strategy. That is a default setting.
We run AI inside real marketing workflows every day, so we think about model selection the way we think about picking a bidding strategy: match the tool to the job, not the hype. Ethan Mollick's opinionated guide to which AI to use made the same point for general knowledge work. Here is our version, scoped to the tasks marketing teams actually ship.
Stop asking "which model is best"
There is no best model. There is only the best model for a task under constraints you care about: quality, cost, latency, and how much of the work is reasoning versus retrieval versus formatting.
Most marketing AI disappointment comes from a category error. Teams pick one frontier model, wire it into everything, and then wonder why their reporting summaries are brilliant but their bulk ad-copy variations cost a fortune and take forever. The model was never the problem. The routing was.
So before you compare models, sort your tasks into three buckets:
- Generation tasks produce text, variations, or drafts. Ad copy, email sequences, meta descriptions, first-draft blog outlines.
- Analysis tasks reason over data or documents. Reading a GA4 export and telling you why conversions dropped, auditing a landing page, synthesising customer research.
- Automation tasks are the glue. Classifying inbound leads, tagging content, routing tickets, transforming one format into another inside a pipeline.
Each bucket rewards different model traits. Let's map them.
Generation: quality per dollar, at volume
Generation is where most marketing spend on tokens goes, because you rarely generate one thing. You generate forty subject lines, twelve ad headlines, and eight variations of a value proposition.
For high-visibility, brand-defining copy, use a frontier model and accept the cost. Homepage messaging, a flagship campaign, a founder's LinkedIn voice. These are low-volume, high-stakes. The marginal token cost is irrelevant next to getting it right.
For high-volume, lower-stakes generation, a mid-tier or fast model wins. Producing 200 product-description variants does not need the smartest model in the world. It needs one that is fast, cheap, and good enough with a strong prompt and examples.
The trap here is using your most expensive model for everything because it "writes best." It does write best, and you will feel that every month on the invoice. Reserve the top tier for the copy a human will actually stare at.
One caveat on quality: heavy AI editing tends to flatten brand voice, because models pull toward the statistically likely phrasing. For generation, always feed real examples of your voice and edit the output. Do not ship raw.
Analysis: reasoning beats speed
Analysis is where the frontier models earn their price. When you ask a model to read a messy GA4 report, cross-reference it with a Google Ads export, and explain why cost-per-acquisition climbed, you are asking for genuine reasoning across noisy inputs. This is not a place to economise.
Use the strongest reasoning model you have access to for:
- Diagnosing performance shifts across data sources
- Auditing landing pages against conversion principles
- Synthesising qualitative research into positioning
- Anything where a wrong-but-confident answer costs you a decision
The key selection criterion for analysis is not raw benchmark scores. It is how the model behaves when it is uncertain. A good analysis model tells you what it cannot see. A bad one invents a plausible cause for your traffic drop and sends you chasing ghosts.
This is also where connecting the model directly to your data matters more than the model choice — the mechanics of which we walked through in MCP in marketing.
Automation: cheap, fast, and boringly reliable
Automation tasks are the opposite of analysis. You want the cheapest model that clears the reliability bar, because these tasks run thousands of times and often run unattended.
Classifying a lead as hot or cold, deciding which content bucket a scraped article belongs in, reformatting a draft into your CMS structure. None of these need a genius. They need consistency and a low per-call cost.
Our selection rule for automation: start with the smallest, fastest model, test it against a fixed set of examples, and only move up a tier if accuracy fails your threshold. Most teams do this backwards. They wire a frontier model into an n8n node, it works, they ship it, and they never notice they are paying premium rates to decide whether an email says "pricing" or "support."
We run n8n agent workflows that harvest source material, draft content, and track what ships. Inside those workflows we deliberately route different steps to different models. The harvesting and classification steps use fast, cheap models. The drafting step, where quality shows, uses a stronger one. Same workflow, three different models, chosen per step. That is the whole idea — and it is the same logic we applied to ad agents when GPT-5.6 split into three pricing tiers.
The Opus 5 wrinkle: what actually moved
One reason to revisit your matrix this summer is that a new model landed on 24 July — and it is worth being precise about what changed, because the headline is easy to misread.
Anthropic describes Claude Opus 5 as coming close to the frontier intelligence of Claude Fable 5 at half the price. That is half of Fable 5's price, not a discount on Opus. As Simon Willison noted the same day, Opus 5 is priced the same as its predecessor, Opus 4.8, and it went straight to the top of the Artificial Analysis leaderboard, ahead of Fable 5 itself.
So the price of a token did not fall. What fell is the price of frontier-grade reasoning. In matrix terms, that moves one boundary — the analysis row. Tasks you had pushed down a tier because the strongest model was too expensive can move back up without breaking the budget. The framework does not change. One line in it does.
The habit matters more than this particular release. Every time a model lands, ask which boundary in your matrix it moves, not whether you should re-platform. The correct routing decision from three months ago can quietly become the wrong one.
A working selection checklist
When a new task lands, we ask five questions in order:
- Is this generation, analysis, or automation? This sets the default tier.
- How visible is the output? A human staring at it justifies a stronger model. A background job does not.
- How often does it run? High frequency pushes you toward cheaper models hard.
- Does it need real data access? If yes, grounding via MCP or an API matters more than the model brand.
- What does a confident wrong answer cost? High stakes means pay for the model that admits uncertainty.
Run any task through those five questions and the model choice usually picks itself. No hype required.
The point is not the model. It is the match.
Models will keep leapfrogging each other. Prices will keep moving. If your strategy is "use whatever won the last benchmark," you will re-platform every quarter and feel productive while spending more than you should.
A task-to-model matrix is durable in a way that no single model is. When a top-tier model gets cheaper to reach, you adjust one boundary. When a new fast model appears, you swap it into your automation tier and test. The framework survives; the specifics rotate.
We build this routing into the workflows we run: cheap models for the glue, strong models for the judgement calls, and real data access wherever a decision depends on numbers. If you want help mapping your marketing tasks to the right models — and wiring Claude into your Google Ads, BigQuery and GTM stack through MCP — that is what our AI solutions work is. We start with what you actually need to decide, not with which model is trending.