Which AI to Use for Advertising: The Real Ranking by Job

Black ink illustration of a three-step podium, each step topped with a stylized AI lab logo shaped like a medal, set against a background of voting ballots.

Every week, some company puts out a press release claiming its model “beats GPT” or “crushes Gemini” on an in-house benchmark. Written by the one who ran the test. Convenient.

There’s a place where the lab doesn’t grade its own homework. Arena.ai (formerly LMArena) puts two AIs head to head, blind, on the same prompt. The user votes for the better answer without knowing which model said what. Millions of votes, a leaderboard per use case. Here are the three that matter for a marketing job, as of July 10, 2026.

Writing and copywriting

  1. Claude Opus 4.6 Thinking (Anthropic), score 1500
  2. Claude Fable 5 (Anthropic), score 1499
  3. Claude Opus 4.7 Thinking (Anthropic), score 1489

A 100% Anthropic podium. Yes, this piece is written with a Claude, the coincidence is obvious. Check it yourself on the Creative Writing leaderboard, it moves every week.

Image generation (ad creative)

  1. GPT Image 2, medium (OpenAI), score 1385
  2. Reve 2.1 (Reve), score 1302
  3. Muse Image (Meta), score 1280

None of the names you see most on LinkedIn (Midjourney, Gemini, Grok) made the podium. Reve, in particular, is an outsider few media traders have tried.

Code and no-code

  1. Claude Fable 5 (Anthropic), score 1649
  2. GPT-5.6 Sol, Codex harness (OpenAI), score 1636
  3. GLM-5.2 Max (Z.ai), score 1580

A Chinese name (Z.ai, formerly Zhipu) in the global top 3 for code, wedged between Anthropic and OpenAI. It never makes the headlines here, it probably should.

What this leaderboard is actually worth

A blind vote measures a preference on one isolated prompt. Not reliability on a real client brief, not price, not how well it plugs into your stack. A model can win the vote and still lose the account.

This leaderboard is dated July 10, 2026. It will have already shifted in a month, maybe in a week. The link is worth more than the table: arena.ai/leaderboard.