How often does an AI recipe break a food restriction? We tested 11 models
We asked 11 AI models for 1,530 family dinners, each for a table with several allergies, diets or conditions, and spelled out every person’s rules in the prompt. 143 first drafts (9%) failed the check: the best model missed 1 dinner in 90, the worst 21.
Plan your dinnersFree to start on iPhone, Android and the web.
Updated November 9, 2026
What we asked
Gluten plans one dinner for a whole family, so before we picked a model to write our recipes, we tested 11 of them on that job. We call the test GlutenBench. Every model got the same 30 households, each with several needs at one table: a parent on AIP with Hashimoto’s next to a pregnant partner and two small kids; celiac, low-FODMAP, type 2 diabetes, GERD, PCOS, vegan and keto; all nine major food allergies; kosher with a toddler, halal with a peanut allergy; alpha-gal, MCAS and Crohn’s; a baby on solids and a 78-year-old.
We asked for three dinners per household, and the prompt spelled out every person’s rules: the foods each restriction rules out and the ones to flag. Then Gluten’s food checker read every ingredient, step and swap in the first draft against everyone at the table.
What we found
143 of the 1,530 first drafts (9%) failed the check. The models were far apart: Gemini 3.1 Flash-Lite missed 1 dinner in 90, and Claude Haiku 5.5 missed 21.
| Model | Made by | First drafts that failed | Share |
|---|---|---|---|
| Claude Haiku 5.5 | Anthropic | 21 of 90 | 23% |
| DeepSeek V4.1 Flash | DeepSeek | 20 of 90 | 22% |
| Nemotron 3.5 Super | NVIDIA | 18 of 90 | 20% |
| Mercury 3 | Inception | 18 of 90 | 20% |
| gpt-oss-120b | OpenAI | 25 of 180 | 14% |
| Gemini 3.5 Flash-Lite | 16 of 180 | 9% | |
| GPT-6 Luna | OpenAI | 10 of 180 | 6% |
| Beam 1.0 | Reflection | 6 of 180 | 3% |
| Gemini 3.6 Flash | 3 of 90 | 3% | |
| Gemini 3.8 Flash | 5 of 270 | 2% | |
| Gemini 3.1 Flash-Lite | 1 of 90 | 1% |
Failing the first draft isn’t the end of a recipe. In Gluten, a draft that fails goes back once for a fix, and the checker drops it if it still fails. But a first draft is what you get when you ask a chatbot for dinner.
The slips, in plain words
Most misses weren’t exotic. They’re the swaps a tired cook makes too. Across all 11 models, the most common were:
- Plain flour, penne or fusilli in a dinner for someone with celiac disease: 10 drafts.
- Raw carrot sticks or coins on a toddler’s plate: 10. The CDC lists pieces of hard raw vegetables as a choking hazard for young children.
- Tamari, a soy sauce, in a low-sodium dinner: 7.
- White rice for someone on AIP, where grains are out: 6. Black pepper (4) and rice vinegar (3) too. Our AIP food list has the rest.
- Black beans for someone on keto: 6.
- Feta for someone dairy-free, and salmon or chicken thighs in a dinner meant for a vegetarian: 3 each.
And when we asked a quick question
We also asked the models 34 “can she eat this?” questions about the same households, twice each. 7 of the 11 gave at least one unsafe answer in 68, after we took out 3 answers our checker misread. Among them: yes to spinach for a toddler on a low-histamine diet, cheese or brie for a parent on AIP, butter in the mash of a toddler with a milk allergy, and breadcrumbs for someone with celiac disease.
What this means for dinner
- A chatbot’s recipe is an idea, not a check. Read every ingredient against everyone you’re cooking for, and the label of anything packaged. Even the best model here missed about 1 dinner in 100, with the rules in front of it.
- Two diets at once is where it slips. A low-sodium dinner that leans on tamari, or an AIP plate with a side of rice: each one is fine for somebody else at the table.
- Kids’ plates need their own look. Raw carrot sticks are a healthy side for a grown-up and a choking risk for a 3-year-old.
That’s why Gluten checks every recipe with a list of foods and rules, not with another model. You can run the same checker on any food, for up to four people, with our free food allergy checker. It reads ingredients, not kitchens, so labels and cross-contact still need your eye.
How we ran it
- When and how: late October 2026, through OpenRouter, only on hosts that keep no data, with the recipe prompt Gluten used at the time. Every call streamed, the same way the app runs them.
- How many: 30 households × 3 dinners = 90 per run. Four models ran twice and Gemini 3.8 Flash three times (once on its fastest hosts), and we pooled each model’s runs: 90 to 270 dinners a model, 1,530 in all.
- What counts as failed: the checker found an ingredient, step or swap that broke a rule for someone at the table, or the model never sent the dinner. GPT-6 Luna and Mercury 3 each had 2 sets of 3 dinners that failed before any recipe arrived, and those count.
- Whose rules: ours. The checker is strict on purpose. On AIP it leaves out black pepper and seed spices, which some AIP lists allow. Another dietitian’s list could count a few dinners differently.
- Small numbers: 90 to 270 dinners per model, so a gap of a few points between neighbours is noise. These are the models as of late October 2026, and newer versions may do better or worse.
- Not shown: how good the dinners were. A judge model from another lab scored flavour, kid appeal and practicality separately. Claude Haiku 5.5, for one, wrote some of the best-scored dinners.
Questions about the method, or want the numbers for a story? Write to support@winding.ai.
Cooking for more than one diet?
Add everyone you cook for, with their allergies, diets and ages. Gluten checks every recipe for each of them and tells you what to change for anyone who can’t eat it as written. It’s free to start.
Plan your dinnersQuestions
How often do AI models get allergy and diet recipes wrong?
In our test of 11 models, 143 of 1,530 first drafts (9%) failed our food check, even with every person’s rules in the prompt. The best model missed 1 dinner in 90 and the worst missed 21.
Which model was the most careful?
For recipes, Gemini 3.1 Flash-Lite (1 of 90) and Gemini 3.8 Flash (5 of 270). Careful at one job wasn’t careful at every job: asked 34 “can she eat this?” questions twice, Gemini 3.1 Flash-Lite gave 5 unsafe answers in 68, and Gemini 3.8 Flash gave none.
What did the models get wrong most often?
Everyday swaps: plain flour or wheat pasta for someone with celiac disease, raw carrot sticks on a toddler’s plate, tamari in a low-sodium dinner, white rice or black pepper for someone on AIP, and black beans for someone on keto.
Should I trust a chatbot’s recipe for a food allergy?
Use it for ideas, then read every ingredient yourself and check the labels before you cook. Even the best model here missed about 1 dinner in 100, and these models were given each person’s rules up front, which a quick chat usually doesn’t do.
How does Gluten check its own recipes?
Gluten’s food checker reads every ingredient, step and swap against each person at the table before a recipe reaches a plan. A recipe that fails goes back once for a fix and is dropped if it still fails. The checker is a list of foods and rules, not a model, so the same food gets the same answer every time.
More guides
Gluten gives food ideas, not medical advice. Recipes are checked against the restrictions you enter, but always check labels for allergens and cross-contact, and talk to your doctor or dietitian about your diet.