HelloFresh vs Blue Apron
Meal kits48 scenarios
Panel, equal weight
- Claude Haiku 5.5anthropic/claude-haiku-5.5
- DeepSeek V4.1 Flashdeepseek/deepseek-v4.1-flash
- Gemini 3.8 Flashgoogle/gemini-3.8-flash
- GLM-5.3 Flashz-ai/glm-5.3-flash
- GPT-6 Lunaopenai/gpt-6-luna
- Grok 4.7x-ai/grok-4.7
- Kimi K3.1 Promoonshotai/kimi-k3.1-pro
- Llama 4 Maverickmeta-llama/llama-4-maverick-17b-128e-instruct
- Mistral Small 2603mistralai/mistral-small-2603
The answer
LikelyFast panel
HelloFresh
HelloFresh 62% · Blue Apron 38% — all 9 models combined
4 clearly prefer HelloFresh · 5 with no clear preference · 0 clearly prefer Blue Apron · excluded: 1 · all 9 counted models are in the overall result
95% interval for HelloFresh: 51–73%
Detailseach model, intervals, conditions
All models combined
Fast panelLikelyAll 9 models combined lean HelloFresh (the interval excludes 50%); on their own: 4 clearly prefer HelloFresh · 5 with no clear preference · 0 clearly prefer Blue Apron.
HelloFresh62%
Blue Apron38%
- Confidence interval for HelloFresh: 51–73%
Consensus95% interval 51–73%
Each model
Share for the brand named in each row, with its interval
Bar: HelloFresh from the left, Blue Apron from the right; dashed line = 50%.
| Judge (API model) | Brand and share | Confidence interval | Leaning | Interval chart |
|---|---|---|---|---|
| Claude Haiku 5.5 | HelloFresh: 58% | 46–70% | No clear preference | |
| DeepSeek V4.1 Flash | HelloFresh: 52% | 38–66% | No clear preference | |
| Gemini 3.8 Flash | HelloFresh: 69% | 55–81% | Clear preference | |
| GLM-5.3 Flash | HelloFresh: 68% | 54–80% | Clear preference | |
| GPT-6 Luna | HelloFresh: 56% | 44–67% | No clear preference | |
| Grok 4.7 | HelloFresh: 70% | 56–82% | Clear preference | |
| Kimi K3.1 Pro | HelloFresh: 58% | 46–69% | No clear preference | |
| Llama 4 Maverick | HelloFresh: 82% | 69–91% | Clear preference | |
| Mistral Small 2603 | HelloFresh: 48% | 34–62% | No clear preference |
Replay it as a dropSimulated
Simulation: scenario paths made up to reproduce the published per-judge shares, not measured rows.
The AI consensus leans to HelloFresh
Published 95% confidence interval for HelloFresh: 51–73%. Press Drop to watch simulated scenarios land.
How to read it
Each ball is a simulated buying scenario. Each row of pins is an AI model, pushing the ball toward the brand it picked.
One drop is one sample. Drop again and the pile shifts; the orange band shows where the balance lands in 95% of drops.
A leader is named only when the band clears the middle. The paths are simulated from the published per-judge shares: no model saw these scenarios.
- Other needs
Tap to drop. Hold to keep dropping.
The numbers behind the machine
9dd1e82e400eJSON
Embed
Free to embed. It shows the latest edition; keep the source line and the link.
How this was measured · verify
How to read
The share is how often the models chose each brand, with its 95% interval. A brand leads only when its whole interval is above 50%, for all models combined and for each model on its own; otherwise it is too close to call.
Each page shows the latest edition. Dates are in the downloadable JSON.
Who was counted
- The pre-registered inclusion rule counted 9 of the 10 judges on the panel.
- The results use 48 of the 48 planned scenarios.
Consensus weights of 9 counted judges
| Judge (API model) | Model family | Weight in the consensus |
|---|---|---|
| Claude Haiku 5.5 | Anthropic | 11.1% |
| DeepSeek V4.1 Flash | DeepSeek | 11.1% |
| Gemini 3.8 Flash | 11.1% | |
| GLM-5.3 Flash | Z.ai | 11.1% |
| GPT-6 Luna | OpenAI | 11.1% |
| Grok 4.7 | xAI | 11.1% |
| Kimi K3.1 Pro | Moonshot AI | 11.1% |
| Llama 4 Maverick | Meta | 11.1% |
| Mistral Small 2603 | Mistral AI | 11.1% |
Judges not counted
jdg-live-qwen3-8-flash-or-v1- Too few valid answers (below the pre-registered minimum)
Need-level results
Need-level results are not published: 6 scenarios per need, below the pre-registered minimum of 30.
8 buyer needs
- Cooking skills and world flavors
- Family-friendly favorites
- Flexible skips and swaps
- High-protein meals
- Lowest cost per serving
- Organic ingredients
- Vegetarian variety
- Fast weeknights, little cleanup
Methodology
Each judge is an AI model called through an API; “About this result” names each model and how it was called. In each buyer scenario, the judge chooses between the two brands. The scenario is the unit of analysis.
- Version
ACM-1.0- Precision
- Measured
- Consensus, needs and languages
- 95% confidence intervals (MOVER combination of the per-judge intervals, widened for small samples in battles and newer boards). The consensus gives equal weight to each model family. The per-judge results in “Each model” are the primary measure.
- Per judge
- 95% confidence intervals (hull of the scenario-clustered bootstrap and Wilson intervals, widened for small samples in battles and newer boards). A single model's view is one judge, not the consensus.
- Panel
- Fast panel: the fast or small current models of each vendor, not the full flagship panel.
Data warnings
- Brand names only: the models saw each option as a brand name, with no product facts, and answered from their own knowledge.
About this result
- AI models (API)
- Claude Haiku 5.5
- DeepSeek V4.1 Flash
- Gemini 3.8 Flash
- GLM-5.3 Flash
- GPT-6 Luna
- Grok 4.7
- Kimi K3.1 Pro
- Llama 4 Maverick
- Mistral Small 2603
Every model was called via OpenRouter (pinned providers).
- Conditions
- Each model chose between the two brands in 48 buyer scenarios. Scenario languages: English.
- Pre-registration (SHA-256)
6406cb28c637…What the pre-registration fixes- Signed snapshot
Verify this result: download the signed snapshot (JSON)
- Signing key (kid)
FnG9tqyrr6Y0…· Production key- SHA-256 of the payload
9dd1e82e400e…
Rank, inclusion and influence on results are never sold.