AI for performance marketing
How Google's Gemini 4 Argon performs on 16 performance-marketing tasks, from ad concepts to campaign readouts, and what it costs per output.
On Benmark's 16 performance-marketing tasks, Gemini 4 Argon is level with Claude Opus 5.5 (6 wins, 6 losses and 4 ties) and a little behind Claude Fable 5.2 (5 wins, 8 losses and 3 ties), mostly by small margins. It costs a fraction of either: their estimated costs per output are a median 13 and 16 times Argon's recorded spend. Argon has the outright top score for adapting a reference ad (8.33), the joint top score for ad analysis (9.93) and the second-highest score for performance readouts (9.14). Its weakest results are narrative development (7.75, against 9.00 at the top), asset naming (8.21 against 9.12) and concepting from a brand brief alone (7.22 against 8.11).
Average across all 16 tasks: 8.72, third of the 12 models tested on every task.
Argon is still the value pick: near the top on analysis and concepts, for cents. Carbon is better on most creative and insight work, but at 2.5 times the cost (both models’ rates double in December), and I'd still hand narrative, briefs and asset naming to Claude.
Analysis and content review: near the top, for cents. Argon has the second-highest analysis average (9.69, after Claude Fable 5.1's 9.75) and the third-highest content-review score (8.00, after Gemini 4 Carbon's 8.46 and Gemini 3.5 Flash's 8.38). It shares the top score for ad analysis (9.93) and is second on performance readouts (9.14). On readouts, ad analysis and content review it costs $0.01 to $0.24 per output.
Concepting: close behind the leaders. Argon has the outright top score for adapting a reference ad (8.33) and the fourth-highest ideation average (8.05, after Gemini 4 Carbon's 8.30, Claude Opus 5.5's 8.23 and Claude Fable 5.2's 8.12). Its concepts cost $0.01 to $0.03 each.
Narrative, briefs and asset naming: use a Claude model. Argon's narrative (7.75) trails Fable 5.2's 9.00. Its compliant brief (9.17) sits below the 9.50 that 12 models reach. Its asset naming (8.21) trails Fable 5.2 (9.12) and Opus 5.5 (8.65). The narrative and brief results each rest on one paired observation.
Scaffold for analysis, simple prompt for proposals. With the scaffold, Argon's performance readouts rise from 8.36 to 9.14 (7 paired observations) and ad analysis from 9.30 to 9.93 (10). The simple prompt did better for proposals from campaign insights: 7.33 against 6.89 (3).
For most marketing work, yes. Argon beats 3.8 Flash on 11 tasks, loses 1 (compliant briefs, 9.17 against 9.50) and ties 4, at a median 3.5 times the cost per output. The biggest gaps are asset naming (8.21 against 5.27), insight codification (7.44 against 5.25) and content review (8.00 against 6.25). For high-volume tagging, Flash is close (9.47 against 9.49) at a little over half the cost.
| Compared with | Gemini 4 Argon ahead | Behind | Level | Gemini 4 Argon’s cost per output |
|---|---|---|---|---|
| Gemini 4 Carbon | 5 | 8 | 3 | 2.5× lower (median of 12 tasks) |
| Claude Fable 5.2 | 5 | 8 | 3 | 16× lower (median of 10 tasks) |
| Claude Opus 5.5 | 6 | 6 | 4 | 13× lower (median of 7 tasks) |
| GPT-6 Astra | 12 | 1 | 3 | 4.4× lower (median of 12 tasks) |
| Gemini 3.8 Flash | 11 | 1 | 4 | 3.5× higher (median of 12 tasks) |
Each row compares published scores task by task, with each model on its better prompt setup; many margins are small. Costs for Gemini 4 Carbon, Gemini 4 Argon, GPT-6 Astra and Gemini 3.8 Flash are recorded spend. Costs for Claude Fable 5.2 and Claude Opus 5.5 are API-equivalent estimates. Gemini 4 Carbon's and Gemini 4 Argon's costs were recorded at their introductory prices, which end on 30 November 2026. Mixed-basis costs are excluded.
Carbon wins 8 tasks, Argon 5, and 3 are level. Argon's wins are all by a third of a point or less, the largest on adapting a reference ad (8.33 against 8.00). Carbon's run up to a point, on insight codification (8.46 against 7.44). Argon costs a median 2.5 times less per output.
Fable 5.2 wins 8 tasks, Argon 5, and 3 are level. Argon's clearest leads are adapting a reference ad (8.33 against 7.94), performance readouts (9.14 against 8.79) and content review (8.00 against 7.67). Fable 5.2's are narrative development (9.00 against 7.75, one paired observation) and asset naming (9.12 against 8.21). Fable 5.2 costs a median 16 times more per output.
Each wins 6 tasks, and 4 are level. Argon leads on hooks (8.25 against 6.75, one paired observation) and adapting a reference ad (8.33 against 7.72). Opus 5.5 leads on concepting from a brand brief (8.11 against 7.22) and asset naming (8.65 against 8.21). Opus 5.5 costs a median 13 times more per output.
Argon wins 12 tasks and loses 1, compliant briefs (9.17 against 9.50), with 3 level. Its biggest leads are content review (8.00 against 5.12, 20 paired observations) and iterating on a top performer (9.33 against 8.03, 10). Astra costs a median 4.4 times more per output.
| Stage | Task | Score | Prompt | Rank | Sample | Gap to top | Cost per output | Status |
|---|---|---|---|---|---|---|---|---|
| Ideation | ||||||||
| Ideation | Reference ad to new concept | 8.33 | Simple | 1 of 18 | 6 paired | Top | $0.031 | Provisional |
| Ideation | Brand context to concepts | 7.22 | Simple | 5 of 18 | 3 paired | −0.89 | $0.012 | Provisional |
| Ideation | Insights and ads to a new proposal | 7.33 | Simple | 4 of 18 | 3 paired | −0.67 | $0.018 | Provisional |
| Ideation | Iterate on a top performer | 9.33 | Simple | 4 of 15 | 10 paired | −0.14 | $0.034 | Provisional |
| Planning | ||||||||
| Planning | Idea scoring and priority plan | 10.00 | Simple | =1 of 19 | 1 paired | Top (joint) | $0.122 | Provisional |
| Briefing | ||||||||
| Briefing | Hook development | 8.25 | Scaffolded | 6 of 21 | 1 paired | −0.50 | Not recorded | Published |
| Briefing | Narrative development | 7.75 | Simple | =11 of 21 | 1 paired | −1.25 | Not recorded | Published |
| Briefing | Compliant brief generation | 9.17 | Simple | =16 of 18 | 1 paired | −0.33 | Not recorded | Published |
| Production | ||||||||
| Production | Content review | 8.00 | Simple | 3 of 18 | 20 paired | −0.46 | $0.012 | Provisional |
| Launch | ||||||||
| Launch | Creative tagging | 9.49 | Combined | 4 of 19 | 20 outputs | −0.24 | $0.046 | Provisional |
| Launch | Asset naming and versioning | 8.21 | Simple | 9 of 18 | 6 paired | −0.91 | Mixed basis | Provisional |
| Launch | Deployment QA checklist | 10.00 | Combined | =1 of 17 | 2 outputs | Top (joint) | $0.245 | Provisional |
| Analysis | ||||||||
| Analysis | Performance readout | 9.14 | Scaffolded | 2 of 20 | 7 paired | −0.22 | $0.244 | Provisional |
| Analysis | Why an ad worked | 9.93 | Scaffolded | =1 of 15 | 10 paired | Top (joint) | $0.035 | Provisional |
| Analysis | Creative diversity and coverage | 10.00 | Scaffolded | =1 of 18 | 2 paired | Top (joint) | $1.928 | Provisional |
| Insight | ||||||||
| Insight | Insight codification | 7.44 | Scaffolded | =7 of 19 | 2 paired | −1.02 | $0.282 | Provisional |
| Stage | Tasks | Gemini 4 Argon average | Rank | Stage leader |
|---|---|---|---|---|
| Ideation | 4 | 8.05 | 4 of 14 | Gemini 4 Carbon (8.30) |
| Planning | 1 | 10.00 | =1 of 19 | 12 models tied (10.00) |
| Briefing | 3 | 8.39 | 10 of 18 | Claude Fable 5.2 (9.00) |
| Production | 1 | 8.00 | 3 of 18 | Gemini 4 Carbon (8.46) |
| Launch | 3 | 9.23 | 7 of 16 | Claude Fable 5.2 (9.52) |
| Analysis | 3 | 9.69 | 2 of 14 | Claude Fable 5.1 (9.75) |
| Insight | 1 | 7.44 | =7 of 19 | Gemini 4 Carbon (8.46) |
Averages are unweighted means of the published task scores. Rank counts only models with a score on every task in the stage. Stages cover the tasks in Full results on this page.
Every model gets the same fixed inputs for an invented brand: reference ads, brand briefs, campaign data or asset records, depending on the task. Each input runs with a simple prompt and with a scaffolded prompt that adds a method and criteria, and a model's published score is its better setup. Creative work is blind-scored against a skilled-human standard; structured tasks are scored automatically against answer keys. Each task page sets out its inputs, prompts and scoring. Gemini 4 Argon was tested in October 2026.
Benmark is an independent benchmark of AI models on performance-marketing work, run by Ben, who co-founded and sold a performance marketing agency. No model provider pays for or reviews these results.
Benmark (2026). Gemini 4 Argon for marketing: benchmark results across 16 tasks. https://benmark.ai/models/gemini-4-argon/
Useful, occasional updates
New benchmark results and lab experiments when there’s something worth sharing. No generic newsletter.
Follow the experiments →Ask Ben anything ↗