Paying for brains you don't need: what Optimizely's Mark-Bench means for your AI budget

Optimizely recently published a LinkedIn highlighting that with the use of their Marketing AI Benchmark project Mark-Bench validating the expected cost for usage. In this article, we break down what this means for customer.

The AI number nobody puts on the slide 

100 marketing tasks a day: $362 on Claude Opus 5.5. $68 on Kimi K3. The quality gap? Under five points. 

That's the headline from Optimizely's new Mark-Bench research. It changes the AI question from "which model is best?" to "which model is good enough, and what am I paying for the difference?" Once AI moves from pilot to production, that's the question that counts.

What Optimizely tested 

Mark-Bench puts AI models through 285 real marketing tasks, each with a brief, source files and a checklist. A task only passes if every item is met. Optimizely pitted the closed-weight models behind Mark today (Claude, Gemini, GPT) against open-weight challengers (Kimi, GLM, DeepSeek) it plans to bring into Mark soon. In short: a preview of the choices coming to your instance.

Quality per dollar 

Kimi K3 delivers 92% of Opus's quality for 19% of the price.

Source: Optimizely Mark-Bench (early version), November 2026. Ratios calculated by Niteco; Argon's are approximate.

• The top is a tie. Argon, Opus and GPT-6.1 Sol are statistically level. Paying for a "winner" buys little. 
• The gap is small. Open-weight Kimi K3 and GLM-5.3 sit within six points of Opus at a fifth of the price. 
• Simple work is nearly free. The cheapest model clears 59% of what Opus does for under two cents.

How cents become six figures 

AI is billed by the token, on everything that goes in and comes out. Briefs, brand guides, product data and multi-step agents all run the meter. Small per-task prices compound fast: 

Based on Mark-Bench average cost per task. Real costs vary with task size and setup.

Model choice is a multiplier on every market, language and agent you add.

Right model, right job 

On summarising an email, Optimizely's Imran Yousuf puts it neatly: "A cent and a half might be more than sufficient." The smart play is matching each task to the cheapest model that does it well.

The catch: you can only route well if you know what each model can do. That's what marketing-specific benchmarks are for. 

What this means for Optimizely customers? 

Optimizely is building a model router, per-instance control over which models are switched on, and model recommendations for the agents you build. Model choice is becoming a commercial decision, not just a technical one. The payoff: 

Your budget goes further. Savings on routine work fund more markets, use cases and tests. 

Scale becomes viable. Catalogue enrichment and localisation stop being too expensive to automate. 

You're less locked in. With the top models tied and open-weight close behind, no vendor is irreplaceable. 

Get ready now. Know your cost per task, define what "good enough" means for each job, decide who owns model choices, and check your data requirements before new models land. 

One caveat: this is an early version of Mark-Bench, so treat small gaps as directional. The direction itself is clear. 

Work with the largest global Optimizely partner