Fable 5.2 launched at the top of general leaderboards such as the Vals Index. On marketing work its lead is narrower, and it costs far more than Gemini 4 Argon, which matches it on many tasks.
AI for performance marketing
Claude Fable 5.2 for marketing
How Anthropic's Claude Fable 5.2 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, Claude Fable 5.2 is roughly level with Google Gemini 4 Argon (8 wins, 5 losses and 3 ties, mostly by small margins) at a median of about 16 times the estimated cost per output. It beats Fable 5.1 on 9 tasks, ties 4 and drops on 3, including campaign performance readouts (9.36 to 8.79). It has the outright top score on two tasks: narrative development (9.00, from one paired observation) and asset naming (9.12, against 8.65 for Gemini 4 Carbon and Claude Opus 5.5).
- 8–5–3
- wins, losses and ties against Gemini 4 Argon
- 16×
- median estimated cost per output against Gemini 4 Argon
- 9 of 16
- tasks improved on Fable 5.1
- 2
- tasks with the outright top score (6 including ties)
Average across all 16 tasks: 8.88, second of the 12 models tested on every task.
Ben’s take
Fable 5.2 has the best briefing scores on the benchmark and is close to the top on concepts, so for creative work it is expensive, but worth testing: at under 60 cents a concept, the cost is negligible next to the ad it builds. For anything that runs at volume, such as tagging, QA or reporting, I'd use Gemini 4 Argon, which matches or beats it for a fraction of the price.
Where to use Claude Fable 5.2
Briefing and ideation: top three on every task. Fable 5.2 ranks in the top three on all seven briefing and ideation tasks. It has the highest briefing average of any model (9.00) and the third-highest ideation average (8.12, after Gemini 4 Carbon's 8.30 and Claude Opus 5.5's 8.23). Its three briefing results each rest on a single paired observation, so treat them as directional.
Readouts, ad analysis, review and tagging: use Gemini 4 Argon. On performance readouts, ad analysis, content review and creative tagging, Argon scores higher and costs 12 to 20 times less per output. Fable 5.1 also beats Fable 5.2 on performance readouts (9.36 against 8.79) and creative tagging (9.73 against 9.43).
Naming and insight work: give it the scaffolded prompt. On the same cases, Fable 5.2's asset-naming score rises from 6.08 with the simple prompt to 9.13 with the scaffold (6 paired observations), and insight codification from 4.49 to 7.83 (2). For performance readouts the simple prompt did better: 8.79 against 8.14 (7).
Is Claude Fable 5.2 worth the cost for marketing work?
It depends on volume. On the 10 tasks with comparable costs for both models, Fable 5.2's estimated cost per scored output is 12 to 33 times Gemini 4 Argon's recorded spend (median 16 times), and Argon matches or beats it on 6 of those 10. For high-volume work such as tagging, content review and reporting, that decides it in Argon's favour. For low-volume creative work the difference is small in absolute terms: Fable 5.2's ideation outputs cost $0.31 to $0.58 each, against $0.01 to $0.03 for Argon.
Against Claude Opus 5.5 the cost is similar: a median of 1.3 times per output, on the 6 tasks where both are estimated the same way.
How Claude Fable 5.2 compares
| Compared with | Claude Fable 5.2 ahead | Behind | Level | Claude Fable 5.2’s cost per output |
|---|---|---|---|---|
| Gemini 4 Carbon | 3 | 8 | 5 | 6.3× higher (median of 10 tasks) |
| Gemini 4 Argon | 8 | 5 | 3 | 16× higher (median of 10 tasks) |
| Claude Opus 5.5 | 6 | 4 | 6 | 1.3× higher (median of 6 tasks) |
| GPT-6 Astra | 11 | 1 | 4 | 3.7× higher (median of 11 tasks) |
| Claude Fable 5.1 | 9 | 3 | 4 | 1.3× higher (median of 10 tasks) |
Each row compares published scores task by task, with each model on its better prompt setup; many margins are small. Costs for Fable 5.2, Fable 5.1 and Opus 5.5 are API-equivalent estimates. Costs for Gemini 4 Carbon, Gemini 4 Argon and GPT-6 Astra are recorded spend. 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.
Claude Fable 5.2 vs Gemini 4 Carbon
Carbon wins 8 tasks, Fable 5.2 wins 3, and 5 are level. 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). Its estimated cost per output is a median 6.3 times Carbon's recorded spend.
Claude Fable 5.2 vs Gemini 4 Argon
Fable 5.2 wins 8 tasks, Argon 5, and 3 are level. Fable 5.2's clearest leads are narrative development (9.00 against 7.75, one paired observation) and asset naming (9.12 against 8.21). Argon's are reference-ad adaptation (8.33 against 7.94), performance readouts (9.14 against 8.79) and content review (8.00 against 7.67). Argon costs a median 16 times less per output.
Claude Fable 5.2 vs Claude Opus 5.5
Fable 5.2 wins 6, Opus 5.5 wins 4, and 6 are level. Fable 5.2 is far ahead on hooks (8.50 against 6.75) and narrative (9.00 against 7.75), both from one paired observation. Opus 5.5 leads on brand-only concepting (8.11 against 7.33). Cost per output is similar.
Claude Fable 5.2 vs GPT-6 Astra
Fable 5.2 wins 11 tasks and loses 1, creative tagging (9.43 against 9.47), with 4 level. Its biggest leads are content review (7.67 against 5.12, 20 paired observations) and turning insights into proposals (7.78 against 6.22). Astra's recorded spend is a median 3.7 times lower.
Is Claude Fable 5.2 better than Fable 5.1?
On most tasks, yes. It improves on 9, matches 4 and drops on 3. The biggest gains are asset naming (6.48 to 9.12) and insight codification (5.62 to 7.83). The drops are performance readouts (9.36 to 8.79), creative tagging (9.73 to 9.43) and why an ad worked (9.90 to 9.77). Estimated cost per output is a median 1.3 times higher.
Full results
| Stage | Task | Score | Prompt | Rank | Sample | Gap to top | Cost per output | Status |
|---|---|---|---|---|---|---|---|---|
| Ideation | ||||||||
| Ideation | Reference ad to new concept | 7.94 | Simple | 3 of 18 | 6 paired | −0.39 | $0.578 | Provisional |
| Ideation | Brand context to concepts | 7.33 | Simple | =3 of 18 | 3 paired | −0.78 | $0.399 | Provisional |
| Ideation | Insights and ads to a new proposal | 7.78 | Simple | 2 of 18 | 3 paired | −0.22 | $0.308 | Provisional |
| Ideation | Iterate on a top performer | 9.43 | Scaffolded | =2 of 15 | 10 paired | −0.04 | $0.513 | Provisional |
| Planning | ||||||||
| Planning | Idea scoring and priority plan | 10.00 | Simple | =1 of 19 | 1 paired | Top (joint) | $2.779 | Provisional |
| Briefing | ||||||||
| Briefing | Hook development | 8.50 | Scaffolded | =2 of 21 | 1 paired | −0.25 | Not recorded | Published |
| Briefing | Narrative development | 9.00 | Simple | 1 of 21 | 1 paired | Top | Not recorded | Published |
| Briefing | Compliant brief generation | 9.50 | Simple | =1 of 18 | 1 paired | Top (joint) | Not recorded | Published |
| Production | ||||||||
| Production | Content review | 7.67 | Simple | 5 of 18 | 20 paired | −0.79 | $0.236 | Provisional |
| Launch | ||||||||
| Launch | Creative tagging | 9.43 | Combined | =8 of 19 | 20 outputs | −0.30 | $0.535 | Provisional |
| Launch | Asset naming and versioning | 9.12 | Scaffolded | 1 of 18 | 6 paired | Top | $4.604 | Provisional |
| Launch | Deployment QA checklist | 10.00 | Combined | =1 of 17 | 2 outputs | Top (joint) | Not recorded | Provisional |
| Analysis | ||||||||
| Analysis | Performance readout | 8.79 | Simple | 6 of 20 | 7 paired | −0.57 | $3.626 | Provisional |
| Analysis | Why an ad worked | 9.77 | Scaffolded | 6 of 15 | 10 paired | −0.16 | $0.467 | Provisional |
| Analysis | Creative diversity and coverage | 10.00 | Simple | =1 of 18 | 2 paired | Top (joint) | Mixed basis | Provisional |
| Insight | ||||||||
| Insight | Insight codification | 7.83 | Scaffolded | 3 of 19 | 2 paired | −0.63 | $3.768 | 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 Claude Fable 5.2’s API-equivalent estimate 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 | Claude Fable 5.2 average | Rank | Stage leader |
|---|---|---|---|---|
| Ideation | 4 | 8.12 | 3 of 14 | Gemini 4 Carbon (8.30) |
| Planning | 1 | 10.00 | =1 of 19 | 12 models tied (10.00) |
| Briefing | 3 | 9.00 | 1 of 18 | Claude Fable 5.2 (9.00) |
| Production | 1 | 7.67 | 5 of 18 | Gemini 4 Carbon (8.46) |
| Launch | 3 | 9.52 | 1 of 16 | Claude Fable 5.2 (9.52) |
| Analysis | 3 | 9.52 | 6 of 14 | Claude Fable 5.1 (9.75) |
| Insight | 1 | 7.83 | 3 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 Claude Fable 5.2
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. Claude Fable 5.2 was tested in November 2026.
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.
- Cost figures use different bases: API-equivalent estimates for the Claude models listed here, recorded spend for Gemini 4 Carbon, Gemini 4 Argon and GPT-6 Astra.
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). Claude Fable 5.2 for marketing: benchmark results across 16 tasks. https://benmark.ai/models/claude-fable-5-2/
Related
- Gemini 4 Carbon
- 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 ↗