Google calls Gemini 4 Carbon its strongest model. Its launch benchmarks cover coding, agents, finance and cybersecurity, with nothing on marketing. It opened to paid API users on 10 November 2026 at the same token prices as Gemini 4 Argon.
AI for performance marketing
Gemini 4 Carbon for marketing
How Google's Gemini 4 Carbon performs on 16 performance-marketing tasks, from ad concepts to campaign readouts, and what it costs per output.
The verdict
On Benmark's 16 performance-marketing tasks, Gemini 4 Carbon is the most consistent model tested. It is the only model tested on all 16 tasks to score 8 or more on every one, and it has the highest average across them (8.96), narrowly ahead of Claude Fable 5.2 (8.88). It beats Gemini 4 Argon on 8 tasks and loses 5, all five by a third of a point or less, but costs a median 2.5 times as much per output at the same token prices. Against Claude Fable 5.2 it wins 8, loses 3 and ties 5, and Fable 5.2's estimated cost per output is a median 6.3 times Carbon's. It has the outright top score on three tasks: turning campaign insights into proposals (8.00), content review (8.46, against Gemini 3.5 Flash's 8.38) and insight codification (8.46).
- 16 of 16
- tasks scored 8 or more, the only model tested on all 16 to do so
- 8.96
- highest average of the 12 models tested on all 16 tasks
- 8–5–3
- wins, losses and ties against Gemini 4 Argon
- 2.5×
- Gemini 4 Argon's cost per output (median), at the same token prices
Average across all 16 tasks: 8.96, first of the 12 models tested on every task.
Ben’s take
Carbon is the first model I've tested that I'd trust across the whole workflow: nothing it does falls below the bar. It isn't as cheap as some others, though. Argon gets you most of the way for about 40% of the cost, so I'd use Carbon where the extra quality matters, on concepts, reviews and insight work, and keep Argon for analysis.
Where to use Gemini 4 Carbon
Concepting and insight work: the highest scores tested. Carbon has the highest ideation average of any model (8.30, ahead of Claude Opus 5.5's 8.23 and Claude Fable 5.2's 8.12). It also has the top scores for turning campaign insights into proposals (8.00) and insight codification (8.46). Those two rest on three and two paired observations, so treat them as directional.
Content review: the top score, narrowly. Carbon scores 8.46 on content review (twenty paired observations), against 8.38 for Gemini 3.5 Flash, at $0.04 per output.
Analysis: Gemini 4 Argon is cheaper and slightly better. Argon scores a little higher on performance readouts (9.14 against 9.07) and ad analysis (9.93 against 9.83). It also costs 2.5 and 1.5 times less per output on those tasks.
Narrative and asset naming: Claude Fable 5.2 is still ahead. Fable 5.2 leads on narrative (9.00 against 8.25, one paired observation each) and asset naming (9.12 against 8.65). It also has the highest briefing average (9.00, against Carbon's 8.75).
Give it the scaffold for analysis and naming. With the scaffold, Carbon's performance readouts rise from 8.14 to 9.07 (seven paired observations), its ad analysis from 9.00 to 9.83 (ten) and its asset naming from 8.09 to 8.65 (six). The simple prompt did better for adapting a reference ad: 8.00 against 7.61 (six).
Is Gemini 4 Carbon worth it over Gemini 4 Argon?
For creative and insight work, yes; for analysis, no. Carbon's token prices match Argon's, but it uses more tokens per task. On the 12 tasks with recorded costs for both, its cost per output is a median 2.5 times Argon's (1.2 to 5.7 times). It beats Argon by up to a point on concepting from a brand brief (8.00 against 7.22), insight proposals (8.00 against 7.33) and insight codification (8.46 against 7.44). On the five tasks Argon wins, it trails by a third of a point or less. Both models' token prices double on 1 December, so the ratio holds.
How Gemini 4 Carbon compares
| Compared with | Gemini 4 Carbon ahead | Behind | Level | Gemini 4 Carbon’s cost per output |
|---|---|---|---|---|
| Gemini 4 Argon | 8 | 5 | 3 | 2.5× higher (median of 12 tasks) |
| Claude Fable 5.2 | 8 | 3 | 5 | 6.3× lower (median of 10 tasks) |
| Claude Opus 5.5 | 8 | 3 | 5 | 5.4× lower (median of 7 tasks) |
| GPT-6 Astra | 11 | 1 | 4 | 1.7× lower (median of 12 tasks) |
| Gemini 3.8 Flash | 11 | 1 | 4 | 5.4× 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.
Gemini 4 Carbon vs Gemini 4 Argon
Carbon wins 8 tasks, Argon 5, and 3 are level. Carbon's biggest leads are insight codification (8.46 against 7.44), concepting from a brand brief (8.00 against 7.22) and insight proposals (8.00 against 7.33). Argon's are adapting a reference ad (8.33 against 8.00) and iterating on a top performer (9.33 against 9.20). Carbon costs a median 2.5 times more per output.
Gemini 4 Carbon vs Claude Fable 5.2
Carbon wins 8, Fable 5.2 wins 3, and 5 are level. Carbon leads on content review (8.46 against 7.67, twenty paired observations), concepting from a brand brief (8.00 against 7.33) and insight codification (8.46 against 7.83). Fable 5.2 leads on narrative (9.00 against 8.25, one paired observation), asset naming (9.12 against 8.65) and iterating on a top performer (9.43 against 9.20). Fable 5.2's estimated cost per output is a median 6.3 times Carbon's recorded spend.
Gemini 4 Carbon vs Claude Opus 5.5
Carbon wins 8, Opus 5.5 wins 3, and 5 are level. Carbon is far ahead on hooks (8.50 against 6.75, one paired observation) and insight codification (8.46 against 7.56). Opus 5.5's three wins are narrow: concepting from a brand brief (8.11 against 8.00), iterating on a top performer (9.43 against 9.20) and ad analysis (9.90 against 9.83). Opus 5.5's estimated cost per output is a median 5.4 times Carbon's recorded spend.
Gemini 4 Carbon vs GPT-6 Astra
Carbon wins 11 tasks and loses one, creative tagging (9.45 against 9.47), with 4 level. Its biggest leads are content review (8.46 against 5.12, twenty paired observations) and insight codification (8.46 against 6.50). Astra's recorded spend is a median 1.7 times Carbon's.
Full results
| Stage | Task | Score | Prompt | Rank | Sample | Gap to top | Cost per output | Status |
|---|---|---|---|---|---|---|---|---|
| Ideation | ||||||||
| Ideation | Reference ad to new concept | 8.00 | Simple | 2 of 18 | 6 paired | −0.33 | $0.170 | Provisional |
| Ideation | Brand context to concepts | 8.00 | Scaffolded | 2 of 18 | 3 paired | −0.11 | $0.069 | Provisional |
| Ideation | Insights and ads to a new proposal | 8.00 | Scaffolded | 1 of 18 | 3 paired | Top | $0.046 | Provisional |
| Ideation | Iterate on a top performer | 9.20 | Scaffolded | 6 of 15 | 10 paired | −0.27 | $0.053 | Provisional |
| Planning | ||||||||
| Planning | Idea scoring and priority plan | 10.00 | Simple | =1 of 19 | 1 paired | Top (joint) | $0.162 | Provisional |
| Briefing | ||||||||
| Briefing | Hook development | 8.50 | Scaffolded | =2 of 21 | 1 paired | −0.25 | Not recorded | Published |
| Briefing | Narrative development | 8.25 | Scaffolded | 4 of 21 | 1 paired | −0.75 | Not recorded | Published |
| Briefing | Compliant brief generation | 9.50 | Simple | =1 of 18 | 1 paired | Top (joint) | Not recorded | Published |
| Production | ||||||||
| Production | Content review | 8.46 | Simple | 1 of 18 | 20 paired | Top | $0.044 | Provisional |
| Launch | ||||||||
| Launch | Creative tagging | 9.45 | Combined | 7 of 19 | 20 outputs | −0.28 | $0.123 | Provisional |
| Launch | Asset naming and versioning | 8.65 | Scaffolded | =2 of 18 | 6 paired | −0.47 | Mixed basis | Provisional |
| Launch | Deployment QA checklist | 10.00 | Combined | =1 of 17 | 2 outputs | Top (joint) | $0.484 | Provisional |
| Analysis | ||||||||
| Analysis | Performance readout | 9.07 | Scaffolded | 3 of 20 | 7 paired | −0.29 | $0.608 | Provisional |
| Analysis | Why an ad worked | 9.83 | Scaffolded | 5 of 15 | 10 paired | −0.10 | $0.052 | Provisional |
| Analysis | Creative diversity and coverage | 10.00 | Simple | =1 of 18 | 2 paired | Top (joint) | $5.394 | Provisional |
| Insight | ||||||||
| Insight | Insight codification | 8.46 | Scaffolded | 1 of 19 | 2 paired | Top | $0.329 | Provisional |
- Scores are out of 10. For human-scored tasks, 8 marks usable work from a skilled practitioner.
- Rank counts every model with a published score on the task; = means joint.
- Prompt is the setup behind the published score; Combined means the score pools both setups.
- Sample is the number of paired observations (the same input run with both prompts), or scored outputs where a complete paired comparison is unavailable.
- Cost is Gemini 4 Carbon’s recorded spend at Google's introductory price per scored output.
- Mixed basis marks a cost recorded on a mixed or unclassified basis; it is left out of all cost comparisons.
- Gap to top is the difference from the highest published score on the task.
- Four tasks are at or near the ceiling, where most models tie: idea scoring, compliant briefs, deployment QA and creative diversity.
By stage
| Stage | Tasks | Gemini 4 Carbon average | Rank | Stage leader |
|---|---|---|---|---|
| Ideation | 4 | 8.30 | 1 of 14 | Gemini 4 Carbon (8.30) |
| Planning | 1 | 10.00 | =1 of 19 | 12 models tied (10.00) |
| Briefing | 3 | 8.75 | 3 of 18 | Claude Fable 5.2 (9.00) |
| Production | 1 | 8.46 | 1 of 18 | Gemini 4 Carbon (8.46) |
| Launch | 3 | 9.37 | 2 of 16 | Claude Fable 5.2 (9.52) |
| Analysis | 3 | 9.63 | 4 of 14 | Claude Fable 5.1 (9.75) |
| Insight | 1 | 8.46 | 1 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.
How we tested Gemini 4 Carbon
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 Carbon was tested on 11 November 2026 at its high reasoning setting.
What these results don’t tell you
- Most results are provisional, and seven rest on very small samples: idea scoring, hooks, narrative, compliant briefs, creative diversity and insight codification (one or two paired observations each) and deployment QA (two scored outputs). Read those as directional.
- Ties at the ceiling say little, and four tasks are there: idea scoring, compliant briefs, deployment QA and creative diversity.
- Scores rate the model's output on fixed inputs. They don't measure live campaign results, and a model can behave differently inside a tool or agent.
- Carbon's costs are recorded spend at Google's introductory price, which ends on 30 November 2026; from 1 December its token prices double. Costs for the Claude models are API-equivalent estimates.
About Benmark
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.
Cite this page
Benmark (2026). Gemini 4 Carbon for marketing: benchmark results across 16 tasks. https://benmark.ai/models/gemini-4-carbon/
Related
- Claude Fable 5.2
- Gemini 4 Argon
- Claude Opus 5.5
- GPT-6 Astra
- Compare every model in the Benchmark Explorer
Useful, occasional updates
Follow the experiments.
New benchmark results and lab experiments when there’s something worth sharing. No generic newsletter.
Follow the experiments →Ask Ben anything ↗