Skip to content
AI stack, job

AI for ad copy: picking a model and testing the output

Last verified

For ad copy, pick a model on three things: how close it gets to your brand voice on a real brief, whether it can read your past winners, and cost per hundred variants. Run the same brief through two models, launch the drafts that pass review, and judge them on cost per result.

The workflow for this job
The workflow for this jobFour steps in a loop: the model proposes, the output is checked, a person approves, then it ships and is measured, and what you learn goes back into the next brief.01 MODEL PROPOSESHook and bodyvariants02 CHECKClaims, policy,brand voice03 PERSON APPROVESPicks whatlaunches04 SHIP AND MEASUREJudge on cost perresultNEXT BRIEFThe workflow for this jobFour steps in a loop: the model proposes, the output is checked, a person approves, then it ships and is measured, and what you learn goes back into the next brief.01 MODEL PROPOSESHook and body variants02 CHECKClaims, policy, brand voice03 PERSON APPROVESPicks what launches04 SHIP AND MEASUREJudge on cost per result
Illustrative flow, simplified.
The model proposes, a check catches what it got wrong, a person approves, and the result ships and is measured. What you learn feeds the next brief.

Choosing a model family for ad copy

Ad copy has no clear leader among model families. Anthropic, OpenAI, Google and xAI each offer a top tier and a cheap tier within a 10x price band, and the vendor pages read on 15 November 2026 give no basis for ranking them. Voice fit, tool access and cost decide the shortlist.

Read the table by the use note and the cost tier, then pick two models to test. For ad copy variants, the listed mid tiers are Claude Sonnet 5.5, GPT-6.1 Sol and GPT-6.1 Terra. For bulk copy and rewrites, the listed cheap tiers are Gemini 3.8 Flash and GPT-6 Luna. Several prices carry end dates (Gemini 4 Carbon on 2026-11-30, Gemini 3.8 Flash on 2026-12-31), so check the row date before you budget.

Models for ad copy: picking a model and testing the output
NameVendorJobStatusPrice as the vendor states itCheckedSource
Claude Fable 5.2claude-fable-5-2.AnthropicAd copy, Market research, Analytics, Paid UA agentsCurrent$8 input / $40 output per MTok.AnthropicAnthropic
Claude Opus 5.5claude-opus-5-5.AnthropicAd copy, Market research, Analytics, Paid UA agentsCurrent$4 input / $20 output per MTok.AnthropicAnthropic
Claude Sonnet 5.5claude-sonnet-5-5.AnthropicAd copy, Market research, Analytics, Paid UA agentsCurrent$2 input / $10 output per MTok.AnthropicAnthropic
Claude Haiku 5.5claude-haiku-5-5.AnthropicAd copy, Market research, Analytics, Paid UA agentsCurrentFrom $0.10 input / from $0.50 output per MTok.AnthropicAnthropic
GPT-6 Astragpt-6-astraOpenAIAd copy, Market research, Analytics, Paid UA agentsCurrent$10 input / $50 output per MTok (standard, short context).OpenAIOpenAI
GPT-6.1 Solgpt-6.1-solOpenAIAd copy, Market research, Analytics, Paid UA agentsCurrent$2 input / $10 output per MTok.OpenAIOpenAI
GPT-6.1 Terragpt-6.1-terraOpenAIAd copy, Market research, AnalyticsCurrent$1 input / $5 output per MTok.OpenAIOpenAI
GPT-6 Lunagpt-6-lunaOpenAIAd copy, Market research, AnalyticsCurrent$0.10 input / $0.50 output per MTok.OpenAIOpenAI
Gemini 4 Carbongemini-4-carbon.GoogleAd copy, Market research, Analytics, Paid UA agentsCurrent$2 in / $10 out per MTok until 2026-11-30, then $4 / $20.Google, stated until 30 November 2026Google
Gemini 4 Argongemini-4-argon.GoogleAd copy, Market research, Analytics, Paid UA agentsCurrent$2 in / $10 out per MTok, rising to $4 / $20 later per page.GoogleGoogle
Gemini 4 Hydrogengemini-4-hydrogen.GoogleAd copy, Market research, Analytics, Paid UA agentsCurrent$1.50 in / $7.50 out per MTok.GoogleGoogle
Gemini 3.8 Flashgemini-3-8-flash.GoogleAd copy, Market research, Analytics, Paid UA agentsCurrent$0.75 in / $3.75 out per MTok until 2026-12-31, then $1.50 / $7.50.Google, stated until 31 December 2026Google
Grok 4.7grok-4.7xAIAd copy, Market research, Analytics, Paid UA agentsCurrent$2 in / $0.50 cached / $6 out per MTok (up to 200k prompt), $4 / $12 above. 500k context, knowledge cutoff May 2026.xAIxAI
Muse Spark 1.1muse-spark-1.1MetaAd copy, Market research, AnalyticsPreviewSee vendordev.meta.ai
Advantage+ creative generative AI features in Ads ManagerMetaStatic creatives, Ad copy, Market researchCurrentSee vendorabout.fb.com
Qwen models (Model Studio)AlibabaAd copy, Market research, Static creatives, AnalyticsCurrentSee vendorAlibaba
JevTypeSafe AIAd copy, Market research, AnalyticsLimited access$42 per billion input tokens ($0.042 per million), input tokens only; output tokens free (currently).TypeSafe AITypeSafe AI

Same-brief test method

A same-brief test sends one brief to two models, reviews the drafts blind, launches the ones that pass and judges both models on cost per result. Keep everything except the model identical, including the brief, the variant count and the settings.

  1. 1.Write one brief

    Use the template below. Include your past winners so you can see whether each model reads and matches them.

  2. 2.Pick two models

    Take two rows from the table, from different vendors or from different tiers of one vendor. Write down the model ID exactly as the vendor writes it.

  3. 3.Generate the same number of variants

    Ask each model for the same count, for example 20 each, one angle per variant.

  4. 4.Review blind

    Strip the model name, shuffle the drafts and score each against the review checks in the table further down. Pass or fail, no half marks.

  5. 5.Launch the passing drafts

    Put the survivors from both models into the same campaign structure with equal budget per variant.

  6. 6.Judge on cost per result

    Compare cost per install, trial or purchase per model once each variant has enough spend. Log the date, model IDs and brief so you can rerun the test later.

Cost per hundred variants is easy to estimate. Multiply the brief's input tokens by the input price and the total output tokens by the output price, each quoted per million tokens, then add the two. Retries and review time come on top. At copy volumes the model fee is usually a small line next to review time and test spend.

Brief template

A model can only match a voice it has seen. Paste real winners, name the claims you can prove, and state what to avoid.

  1. 1.Product and offer

    State what is sold, to whom, and what you can prove.

    PRODUCT: one line
    AUDIENCE: who, country, stage
    OFFER: price, trial, promise you can prove
  2. 2.Voice and winners

    Give the model your own copy to read, with results beside it.

    VOICE: three adjectives
    WINNERS: paste a few past top ads with cost per result
    AVOID: banned words, claims you cannot support
  3. 3.Format and output

    Set the placement and the shape of the answer so drafts compare cleanly.

    FORMAT: platform and placement, primary text max characters, headline max characters
    OUTPUT: variant count, one angle each, plain text, numbered

Review checks before launch

Run the same checks on drafts from both models. A draft that fails one check is rewritten or dropped, whichever model wrote it.

Review checks for AI-written ad copy
CheckWhat to look forIf it fails
ClaimsResults, savings, medical or income promises, and 'best' or 'only' wording that you cannot support.Remove the claim or replace it with one you can document.
Invented numbersStatistics, user counts, ratings or percentages the model added on its own.Delete them. Use only figures from your own data.
Platform policyPersonal attributes, before and after framing, restricted categories, misleading urgency.Rewrite against the current ad policy of the platform you buy on.
Brand voiceBanned words, tone drift, phrases that read like generic marketing copy.Add the failing example to AVOID in the brief and regenerate.
Angle spreadVariants that repeat one hook with different wording.Ask for one distinct angle per variant and drop duplicates.
DisclosureMeta labels ads made with its own gen AI tools as 'AI info' (policy updated 2026-06-01). Rules for third-party AI output were not verified.Read the current platform rules before launch if you use AI output.

Run each draft through the same typed checks before launch: hook, offer, call to action, audience and policy flags.

Check your ad copy with the free ad checker

Skills for copy

A skill is a packaged set of instructions, sometimes with scripts, that an AI client loads for a task. Treat every third-party skill as unreviewed code, and check the licence in each repository before you use or host it. The pack below includes a copywriting skill and an ad-creative skill (read 2026-11-15) and carries a flag for unsourced benchmarks, so do not copy its numbers into your briefs.

Skills for this job
NameJobLicenceChecked
coreyhaines31/marketingskillsAd copy, Landing pages, Analytics, SEO and GEOMITClass A

What this page does not claim

  • It does not name a top model. The table comes from vendor pages read on 15 November 2026, and no output-quality test is shown here.
  • It does not say one model writes better copy than another. That depends on your brief, your voice and your results.
  • Vendor-run claims of higher CTR for one model were not verified and are not used.
  • Prices and limits change. Rows carry a checked date, and dated price changes are noted where the vendor page lists them.
  • Meta's Advantage+ creative features in Ads Manager are listed, but the current 2026 feature list and the price were not verified.
  • Third-party skills are listed for reference, not as reviewed code.

Questions

How do Claude and ChatGPT compare for Meta ad copy?
The sources read give no verified ranking of the two for Meta copy. The listed mid tiers, Claude Sonnet 5.5 and GPT-6.1 Sol, are both priced at $2 input and $10 output per million tokens. Run one brief through both, review blind and compare cost per result.
How do you choose an AI model for ad copy?
Three things decide it: how close the model gets to your brand voice on a real brief, whether it can read your past winners, and cost per hundred variants. Shortlist two models from the table and let launch results settle the choice.
Can a cheap model handle ad copy?
Cheap tiers such as Gemini 3.8 Flash and GPT-6 Luna are listed for bulk copy and rewrites. Test them against a mid tier on the same brief. The fee gap per hundred variants can be small, so quality on your brief usually decides.
Is Meta's built-in copy generator worth using?
Advantage+ creative in Ads Manager offers headline and primary text variations. Its current 2026 feature list and price were not verified. Meta labels ads made with its own gen AI tools as 'AI info' under a policy updated 2026-06-01.

Sources

Each source shows the date it was last checked.

Show all 20 sources
  1. about.fb.com about.fb.com/news/2025/02/gen-ai-transparency-metas-ads-products/
  2. Meta ai.meta.com/blog/muse-spark-1-1/
  3. Google aistudio.google.com/docs/models
  4. Google aistudio.google.com/docs/pricing
  5. Alibaba alibabacloud.com/help/en/model-studio/models
  6. Anthropic claude.com/pricing
  7. dev.meta.ai dev.meta.ai/resources/blog/build-with-muse-spark/
  8. OpenAI developers.openai.com/api/docs/models
  9. OpenAI developers.openai.com/api/docs/pricing
  10. TypeSafe AI docs.typesafe.ai/api/system-one.md
  11. TypeSafe AI docs.typesafe.ai/introduction/quickstart.md
  12. TypeSafe AI docs.typesafe.ai/models.md
  13. xAI docs.x.ai/docs/models
  14. Meta facebook.com/business/news/Introducing-Enhanced-Gen-AI-Features-and-Other-Tools-to-Help-Build-Your-Business
  15. learn.chatgpt.com learn.chatgpt.com/docs/pricing
  16. GitHub coreyhaines31/marketingskills: LICENSE
  17. Anthropic platform.claude.com/docs/en/about-claude/pricing
  18. Anthropic platform.claude.com/docs/en/about-claude/models/overview
  19. GitHub coreyhaines31/marketingskills
  20. TypeSafe AI typesafe.ai

Want this applied to your own ad accounts? Start with a short audit.

See the service