GPT-6 Astra held the top score on the Epoch Capabilities Index (169) until Claude Fable 5.2 passed it. Marketing work is a different test.
AI for performance marketing
GPT-6 Astra for marketing
How OpenAI's GPT-6 Astra 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, GPT-6 Astra trails the other frontier models. It loses 11 tasks each to Claude Fable 5.2 and Gemini 4 Carbon, 10 to Claude Opus 5.5 and 12 to Gemini 4 Argon, while costing a median 4.4 times Argon's recorded spend per output. It has the lowest ideation average of the 14 models with all 4 ideation tasks (6.91), and scores 5.12 on content review, against 8.46 at the top. GPT-6 Sol is level with it (5 wins, 5 losses and 6 ties) at a median 6.3 times lower cost.
- 1–12–3
- wins, losses and ties against Gemini 4 Argon
- 4.4×
- higher median cost per output than Gemini 4 Argon
- 6.91
- ideation average, the lowest of 14 models
- 5–5–6
- against GPT-6 Sol, which costs a median 6.3× less
Average across all 16 tasks: 8.11, tenth of the 12 models tested on every task.
Ben’s take
I wouldn't use GPT-6 Astra for marketing work. Gemini 4 Argon beats it on almost every task for less money, and if you want an OpenAI model, GPT-6 Sol is level with it for a fraction of the price.
Where to use GPT-6 Astra
Concepting and content review: avoid. Astra has the lowest ideation average of the 14 models with all 4 ideation tasks (6.91), and it comes last of 15 for iterating on a top performer (8.03). It scores 5.12 on content review (20 paired observations), against 8.00 for Gemini 4 Argon.
Tagging and hooks: competitive, on small samples. Its creative tagging (9.47, from 2 scored outputs) is joint 5th of 19, and its hooks (8.00, one paired observation) joint 7th of 21.
Scaffold for readouts, ad analysis and proposals. Astra's performance readouts rise from 7.50 to 8.64 with the scaffold (7 paired observations), its ad analysis from 8.67 to 9.33 (10) and its campaign-insight proposals from 5.33 to 6.22 (3). The simple prompt did better for concepting from a reference ad (7.28 against 5.50, 6 paired observations) and from a brand brief (6.11 against 5.89, 3).
Is GPT-6 Astra worth the cost for marketing work?
In these tests, no. Gemini 4 Argon beats it on 12 of 16 tasks and costs a median 4.4 times less per output, both on recorded spend. If you want an OpenAI model, GPT-6 Sol is level with Astra at a median 6.3 times lower cost: Sol for concepting, Astra for analysis.
How GPT-6 Astra compares
| Compared with | GPT-6 Astra ahead | Behind | Level | GPT-6 Astra’s cost per output |
|---|---|---|---|---|
| Gemini 4 Carbon | 1 | 11 | 4 | 1.7× higher (median of 12 tasks) |
| Claude Fable 5.2 | 1 | 11 | 4 | 3.7× lower (median of 11 tasks) |
| Gemini 4 Argon | 1 | 12 | 3 | 4.4× higher (median of 12 tasks) |
| Claude Opus 5.5 | 2 | 10 | 4 | 3.6× lower (median of 7 tasks) |
| GPT-6 Sol | 5 | 5 | 6 | 6.3× 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 GPT-6 Astra, Gemini 4 Carbon, Gemini 4 Argon and GPT-6 Sol 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.
GPT-6 Astra vs Gemini 4 Carbon
Carbon wins 11 tasks. Astra wins one, creative tagging (9.47 against 9.45), and 4 are level. Astra's recorded spend is a median 1.7 times Carbon's.
GPT-6 Astra vs Claude Fable 5.2
Fable 5.2 wins 11 tasks. Astra wins 1, creative tagging (9.47 against 9.43), and 4 are level. The biggest gaps are content review (5.12 against 7.67, 20 paired observations) and campaign-insight proposals (6.22 against 7.78). Fable 5.2 costs a median 3.7 times more per output.
GPT-6 Astra vs Gemini 4 Argon
Argon wins 12 tasks. Astra wins 1, compliant briefs (9.50 against 9.17, one paired observation), and 3 are level. The biggest gaps are content review (5.12 against 8.00) and iterating on a top performer (8.03 against 9.33, 10 paired observations). Astra costs a median 4.4 times more per output.
GPT-6 Astra vs Claude Opus 5.5
Opus 5.5 wins 10 tasks. Astra wins hooks (8.00 against 6.75, one paired observation) and creative tagging (9.47 against 9.43), and 4 are level. Opus 5.5 costs a median 3.6 times more per output.
GPT-6 Astra vs GPT-6 Sol
Each wins 5 tasks, and 6 are level. Astra leads on ad analysis (9.33 against 8.00), hooks (8.00 against 6.75) and performance readouts (8.64 against 7.43). Sol leads on all 4 ideation tasks, by up to 0.67. Astra costs a median 6.3 times more per output.
Full results
| Stage | Task | Score | Prompt | Rank | Sample | Gap to top | Cost per output | Status |
|---|---|---|---|---|---|---|---|---|
| Ideation | ||||||||
| Ideation | Reference ad to new concept | 7.28 | Simple | 15 of 18 | 6 paired | −1.05 | $0.199 | Provisional |
| Ideation | Brand context to concepts | 6.11 | Simple | 16 of 18 | 3 paired | −2.00 | $0.089 | Provisional |
| Ideation | Insights and ads to a new proposal | 6.22 | Scaffolded | 17 of 18 | 3 paired | −1.78 | $0.122 | Provisional |
| Ideation | Iterate on a top performer | 8.03 | Scaffolded | 15 of 15 | 10 paired | −1.44 | $0.147 | Provisional |
| Planning | ||||||||
| Planning | Idea scoring and priority plan | 10.00 | Simple | =1 of 19 | 1 paired | Top (joint) | $0.577 | Provisional |
| Briefing | ||||||||
| Briefing | Hook development | 8.00 | Scaffolded | =7 of 21 | 1 paired | −0.75 | Not recorded | Published |
| Briefing | Narrative development | 7.50 | Scaffolded | =17 of 21 | 1 paired | −1.50 | Not recorded | Published |
| Briefing | Compliant brief generation | 9.50 | Scaffolded | =1 of 18 | 1 paired | Top (joint) | Not recorded | Published |
| Production | ||||||||
| Production | Content review | 5.12 | Scaffolded | =16 of 18 | 20 paired | −3.34 | $0.067 | Provisional |
| Launch | ||||||||
| Launch | Creative tagging | 9.47 | Combined | =5 of 19 | 2 outputs | −0.26 | $0.075 | Provisional |
| Launch | Asset naming and versioning | 8.08 | Simple | =10 of 18 | 1 paired | −1.04 | $0.752 | Provisional |
| Launch | Deployment QA checklist | 10.00 | Combined | =1 of 17 | 2 outputs | Top (joint) | $0.999 | Provisional |
| Analysis | ||||||||
| Analysis | Performance readout | 8.64 | Scaffolded | =7 of 20 | 7 paired | −0.72 | $0.977 | Provisional |
| Analysis | Why an ad worked | 9.33 | Scaffolded | 12 of 15 | 10 paired | −0.60 | $0.154 | Provisional |
| Analysis | Creative diversity and coverage | 10.00 | Scaffolded | =1 of 18 | 2 paired | Top (joint) | $3.129 | Provisional |
| Insight | ||||||||
| Insight | Insight codification | 6.50 | Scaffolded | 14 of 19 | 1 paired | −1.96 | $0.595 | 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 GPT-6 Astra’s recorded spend 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 | GPT-6 Astra average | Rank | Stage leader |
|---|---|---|---|---|
| Ideation | 4 | 6.91 | 14 of 14 | Gemini 4 Carbon (8.30) |
| Planning | 1 | 10.00 | =1 of 19 | 12 models tied (10.00) |
| Briefing | 3 | 8.33 | 11 of 18 | Claude Fable 5.2 (9.00) |
| Production | 1 | 5.12 | =16 of 18 | Gemini 4 Carbon (8.46) |
| Launch | 3 | 9.18 | 9 of 16 | Claude Fable 5.2 (9.52) |
| Analysis | 3 | 9.32 | 8 of 14 | Claude Fable 5.1 (9.75) |
| Insight | 1 | 6.50 | 14 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 GPT-6 Astra
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. GPT-6 Astra was tested in September 2026 and October 2026.
What these results don’t tell you
- Most results are provisional, and nine rest on very small samples: idea scoring, hooks, narrative, compliant briefs, asset naming, creative diversity and insight codification (one or two paired observations each) and creative tagging (two scored outputs) 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: recorded spend for GPT-6 Astra, Gemini 4 Carbon, Gemini 4 Argon and GPT-6 Sol, and API-equivalent estimates for the Claude models.
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). GPT-6 Astra for marketing: benchmark results across 16 tasks. https://benmark.ai/models/gpt-6-astra/
Related
- Gemini 4 Carbon
- Claude Fable 5.2
- Gemini 4 Argon
- Claude Opus 5.5
- 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 ↗