Estimated time savings
benmark score over time
Oct 13, 2026Google Gemini 4 Argon · high73.1% · +0.8 pts
Rebuilt from today's tests and time estimates, using the models available at each date.
A benchmark for paid social teams
73.1% of estimated human time can be saved using AI.
The rest still needs a human.
I've tested AI on 18 tasks across paid social. See which models I'd use for each job, what they cost, and where they still need a human.
Still in beta
653 scored run records · 18 creative tasks tracked · updated Oct 23, 2026
How runs are counted
Explore the full benchmark ↗
Free to explore. Get occasional updates.
See the step-by-step view ↓
01 · Model rankings
Which AI model is best at paid social work?
Average score across up to 16 text tasks, each weighted equally and blind-scored out of 10. A model needs 10 scored tasks to be ranked.
See all 16 ranked models, with scores by stage →
Select a model for its scores on every task
Treat small gaps as level. Cost is the average cost per scored output on the same 6 tasks for every model; 2 models’ costs are estimates at API prices, the rest are recorded spend.
02 · Best model by job
The best model changes with the job.
Pick by task, not by overall rank. On 4 jobs, the best model beats the worst by 3 points or more out of 10.
What does the gap mean?
The gap between the best and worst model on each task, out of 10. A bigger gap means your choice of model matters more.
3.0+ pts · large gap1.5–under 3.0 pts · medium gapUnder 1.5 pts · small gap
Cheap and best
Gemini 3.5 Flash ranks 10th overall, but wins Content review and Insight codification outright, at $0.011 or less per output.
Great at some, poor at others
Claude Fable 5.1 wins Creative tagging and Performance readout, then finishes second-last on Asset naming and versioning and Insight codification.
Ideation
AI can take on 90% of the work
Reference ad to new concept
Take a reference ad and turn what makes it interesting into a new concept for a different brand.
Leading modelGoogle Gemini 4 ArgonScore8.33/10Best–worst gap8.33 best − 6.92 worst1.4 ptsCompare modelsBrand context to concepts
Come up with ad concepts using just the brand, product and audience brief.
Leading modelClaude Opus 5.5Score8.11/10Best–worst gap8.11 best − 5.78 worst2.3 ptsCompare modelsInsights and ads to a new proposal
Use what we've learnt from a campaign, alongside its existing ads, to suggest what to test next.
Leading modelClaude Opus 5.5Score7.67/10Best–worst gap7.67 best − 6.11 worst1.6 ptsCompare modelsIterate on a top performer
Suggest a new version of a successful ad. We score the recommendation, not a finished ad or its future performance.
Leading modelClaude Opus 5Score9.47/10Best–worst gap9.47 best − 8.03 worst1.4 ptsCompare models
Planning
AI can take on 89% of the work
Briefing
AI can take on 95% of the work
Hook development
Turn an ad concept into opening hooks that catch attention and give someone a reason to keep watching.
Leading modelMeta Muse Spark 1.3Score8.75/10Best–worst gap8.75 best − 6.29 worst2.5 ptsCompare modelsNarrative development
Build the story of an ad from a concept and hook, keeping the product and audience in mind.
Leading modelClaude Opus 4.7Score8.58/10Best–worst gap8.58 best − 6.50 worst2.1 ptsCompare modelsCompliant brief generation
Turn an agreed concept into a production brief that follows the brand rules and tells the team what to make.
Joint leaders10 models tiedJoint best9.50/10Best–worst gap9.50 best − 8.81 worst0.7 ptsCompare models
Production
AI can take on 52% of the work
Static image generation
Make a basic static ad from a brief and product reference. We score the image as generated, before any editing.
Leading modelGPT Image 2.5 SunburstScore8.84/10Best–worst gap8.84 best − 7.12 worst1.7 ptsCompare modelsVideo generation
Make a basic eight-second vertical video ad from a brief. We score the full video as generated, before any editing.
Leading modelGemini Omni FlashScore6.68/10Best–worst gap6.68 best − 4.56 worst2.1 ptsCompare modelsContent review
Review 10 images and 10 videos and spot the errors we've put in them, without being told what is wrong.
Leading modelGemini 3.5 FlashScore8.38/10Best–worst gap8.38 best − 4.96 worst3.4 ptsCompare models
Launch
AI can take on 72% of the work
Creative tagging
Label the features of image and video ads using a fixed set of creative categories.
Leading modelClaude Fable 5.1Score9.73/10Best–worst gap9.73 best − 8.97 worst0.8 ptsCompare modelsAsset naming and versioning
Propose clear file names and version numbers for a set of assets, spotting duplicates and clashes along the way.
Leading modelClaude Opus 5.5Score8.65/10Best–worst gap8.65 best − 5.27 worst3.4 ptsCompare modelsDeployment QA checklist
Check a campaign launch pack for missing information and mistakes, then say whether it is ready to go live.
Joint leaders14 models tiedJoint best10.00/10Best–worst gap10.00 best − 9.99 worst<0.1 ptsCompare models
Analysis
AI can take on 76% of the work
Performance readout
Read campaign performance data, explain what has changed and suggest what to do about it.
Leading modelClaude Fable 5.1Score9.36/10Best–worst gap9.36 best − 6.00 worst3.4 ptsCompare modelsWhy an ad worked
Look at an ad and explain why it may have worked. Suggestions for the next version are scored separately.
Joint leadersGoogle Gemini 4 Argon and Claude Opus 5Joint best9.93/10Best–worst gap9.93 best − 8.00 worst1.9 ptsCompare modelsCreative diversity and coverage
Look across an ad portfolio to show what is repeated, what is missing and where a new creative test could help.
All models tiedAll 16 models testedJoint best10.00/10Best–worst gap10.00 best − 10.00 worst0.0 ptsCompare models
Insight
AI can take on 28% of the work
03 · Better prompts
A better prompt clearly helps about 4 times in 10.
On 17 of the 18 jobs, I've run the same model on the same input twice: once with a simple prompt, once with a scaffold, the same prompt plus frameworks, guardrails and practical guidance. Across 203 paired comparisons:
clearly helped
79about 4 in 10
little or no change
91about 4 in 10
clearly hurt
33about 2 in 10
“Clearly” means a change of 0.25 points or more out of 10. Smaller moves are treated as no change.
Average change from adding a scaffold, by job
Each model weighted equally · change out of 10
- Insight codification+1.359 better · 1 worse
- Creative diversity and coverage+0.836 better · 0 worse · 10 already 10/10
- Video generation+0.522 better · 0 worse
- Why an ad worked+0.528 better · 0 worse
- Asset naming and versioning+0.324 better · 2 worse
- Performance readout+0.295 better · 2 worse
- Compliant brief generation+0.287 better · 1 worse
- Hook development+0.2610 better · 3 worse
- Idea scoring and priority plan+0.265 better · 0 worse · 5 already 10/10
- Static image generation+0.172 better · 0 worse
- Iterate on a top performer+0.133 better · 0 worse
- Content review+0.055 better · 3 worse
- Narrative development+0.057 better · 4 worse
- Brand context to concepts+0.014 better · 2 worse
- Deployment QA checklist+0.000 better · 0 worse · 6 already 10/10
- Insights and ads to a new proposal−0.152 better · 6 worse
- Reference ad to new concept−0.710 better · 9 worse
No paired comparisons yet: Creative tagging. Some results rest on one or two observations per model.
The step-by-step view
How much of the work AI can take on.
Each percentage is the share of the estimated workload at that step that AI can take on from humans. Put simply: the amount of human time that can be saved.
0190%
Ideation
Coming up with strong concepts.
Based on 4 tasks0289%
Planning
Choosing what deserves making.
Based on 1 task0395%
Briefing
Turning the concept into a clear plan.
Based on 3 tasks0452%
Production
Creating and reviewing the assets.
Based on 3 tasks0572%
Launch
Preparing the work for market.
Based on 3 tasks0676%
Analysis
Understanding what happened.
Based on 3 tasks0728%
Insight
Carrying the learning forward.
Based on 1 taskThe averages hide a bigger story. On some tasks the gap between frontier models is huge. Just because it scores the highest on coding benchmarks, it doesn't mean it is the best for performance marketing.
Get the data
Find the right AI for the job.
Which model should I use?
Compare models by quality and cost, with a winner for each task.
Where could I put it to work?
Explore the scores, estimated time savings and my take on all 18 tasks.
Does a better setup help?
See where extra guidance improves the result and where it adds little.
Methodology
How we get to the number.
Every task follows a declared test and scoring method. The number is built from actual AI runs, human judgement and objective checks designed by a human expert.
- 01Map 18 tasks across the selected creative journey.
- 02Run each task through leading AI models two ways: with a plain prompt (the baseline) and with a purpose-built prompt and workflow (the scaffold).
- 03Score every output blind, with model names hidden: judgement-led work is assessed by a human; objective work is checked against human-designed answer keys and tests.
- 04Estimate the work still needed on each output: individual fixes, human checking and, where necessary, a complete rebuild. A high average cannot cancel out a serious failure.
- 05Compare that work with estimated manual task time. Missing repair evidence receives no saving. Combine the best tested model and prompt approach for each task; business approval gates remain separate.
The person behind benmark
Hi, I'm Ben Nancarrow.
I've spent 15+ years in paid social, built and sold one of the three largest independent media agencies on Meta in EMEA, and worked with 300+ businesses across more than $1bn in ad spend.
Then AI turned up and I got curious. A couple of years of testing later, that curiosity has become benmark. I set the tests, blind-score the creative work and keep asking: “Yes, but is it actually useful?” Apparently, this is what I do for fun.