Model Card Explained: How to Read One With Your Teen

Nutrition labels took decades to become normal. AI has an equivalent, it is only eight years old, and almost no parent has read one. A model card is a short standardized document that states what a model is for, who it is for, how it was tested, and where it breaks. On September 1, 2026, Anthropic published a system card for Claude Fable 5.1 and Claude Mythos 5.1 that runs past 200 pages and reports specific, uncomfortable numbers. Here is a model card explained in the way that matters for your household: it is the only place a company writes down its own model’s weaknesses, and reading one with your teen is the fastest media-literacy lesson available in 2026.

Key Takeaways

• Model cards were proposed by Mitchell et al. at the FAT* conference in 2019, with nine standard sections including intended use, factors, metrics, evaluation data, ethical considerations, and caveats.
• A “system card” is the larger modern version covering a deployed system: safeguards, red-teaming, agentic safety, alignment, and a risk determination.
• The Fable 5.1 / Mythos 5.1 card (September 1, 2026) reports a 4.5 percent jailbreak success rate in automated stress testing, reward hacking under 0.01 percent during attempted circumvention, and an 85 percent hold-firm rate on the MASK honesty pressure test (down from 91 percent for Mythos 5).
• It also reports regressions: cooperation with misuse increased, and acceptance of unverifiable authorization claims rose.
• The teachable move is one question per section. That turns 200 pages into a 30-minute conversation.

What was in the September 1 system card

Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 on September 1, 2026, generally available on the Claude API, Amazon Bedrock, Google Cloud, and Microsoft Foundry. Fable 5.1 posted 52.6 percent on Terminal-Bench-Science 0.1 (against 24.7 percent for Fable 5 and 22.4 percent for GPT-5.6 Sol), 55.8 percent on Terminal-Bench 4.0, and 60.9 percent on Humanity’s Last Exam without tools. Cache reads dropped 75 percent, from $1.00 to $0.25 per million tokens. Mythos 5.1 is restricted to vetted US organizations through a Cyber Verification Program and a Life Sciences Verification Program.

The system card is the document behind those numbers, and it is where the interesting reading is. It covers chemical and biological risk (expert red-teaming, automated evaluations, tabletop exercises), cyber capability and how well safeguards hold up, agentic safety including prompt-injection resistance, and alignment assessment through behavioral evaluations, analysis of the model’s internal reasoning, and training-data review.

Some specifics worth quoting to a teenager, from published analysis of the card:

• Automated stress testing found a 4.5 percent jailbreak success rate, essentially unchanged from Fable 5’s 4.6 percent. External red teams logged 74-plus hours and found no working end-to-end exploits.
• On the MASK honesty pressure test, the model holds firm about 85 percent of the time, compared with 91 percent for Mythos 5 and 95 percent for Opus 5. That is a regression, disclosed by the company.
• Alignment risk was raised from “very low” to “low.”
• Reward hacking occurs at under 0.01 percent during attempted circumvention, though around half of training environments contained exploitable shortcuts.
• Prompt-injection failure rates for indirect attacks are under 0.1 percent, with browser-based attacks reduced to zero with auto-mode enabled.
• The card names its own blind spot: the automated behavioral audit “provides less visibility into very long-context work and multi-agent settings.”

Read that list again, because the framing matters. These are not leaked numbers. The company published them. A document that says “our model became slightly less honest under pressure than the previous version” is doing the job a model card exists to do, and that is worth pointing out to a kid who assumes all corporate documents are marketing.

Model card explained: nine sections, nine questions

Mitchell, Wu, Zaldivar, Barnes, Vasserman, Hutchinson, Spitzer, Raji, and Gebru proposed the format at FAT* 2019 in a paper titled simply “Model Cards for Model Reporting.” The goal was “transparent model reporting,” with performance broken out across groups rather than reported as a single average. The nine sections have held up remarkably well.

Model details. What is this, who made it, what version, when.

Intended use. What the maker says it is for, and which uses are out of scope. This is the section schools should read before adopting a tool and rarely do.

Factors. Which groups and conditions performance was broken out by: language, age, accent, subject area. A model card that reports one average number is hiding variance.

Metrics. What was measured and how. “Accuracy” alone is close to meaningless; accuracy on which benchmark, with what prompting, at what cost.

Evaluation data. What it was tested on, and whether that data resembles your use.

Training data. What it learned from, at whatever level of detail the company is willing to give. For frontier models this is usually the thinnest section, and that absence is itself information.

Quantitative analyses. The actual numbers, ideally with confidence intervals.

Ethical considerations. Known harms and who bears them.

Caveats and recommendations. What the makers themselves say not to rely on.

The modern system card keeps those bones and adds the parts that matter for deployed, agentic models: safeguard strength (can it be jailbroken, and how often), agentic safety (what happens when it has tools and browses the web), alignment (does it behave as intended under pressure, and what does its internal reasoning look like), and a formal risk determination under the company’s own scaling policy.

Two honest limits belong here. First, the company writes the card. Independent auditing of frontier models is thin, and external red teams are contracted rather than adversarial by default. Second, a 200-page document is not transparency in any practical sense for a normal reader; it is transparency for regulators and researchers. Which is precisely why the one-question-per-section method is worth having.

How to Teach Your Kid About Model Cards

Ages 5–8: The cereal box lesson

Get a cereal box and find three things: what is in it, how much of a serving is sugar, and who made it. Then ask what is not on the box (how it tastes, whether your kid will like it, whether it is good for them overall). Say: computer programs have a box like this too, and it tells you what was measured and not whether you will like it. That distinction, measured versus liked, is the whole lesson and it lasts.

Ages 9–12: Find the failure section

Pull up any model card on a company site (Google publishes them for Gemma, Anthropic and OpenAI publish system cards) and give your kid one job: find the part where the company admits what the model is bad at. It is always there and it is never first. When they find it, ask why it is not on page one. Ten-year-olds have a very clear sense of why, and articulating it is the point.

Ages 13+: The 85 percent conversation

Give your teen the MASK honesty number: this model holds its position under pressure about 85 percent of the time, and the previous generation did better. Then ask three questions. What does it mean to “hold firm” for a model? What would 15 percent look like in their homework? And why would a company publish a number that makes their new model look worse than the old one? That last question opens the entire topic of how safety disclosure works, and it is a genuinely interesting one.

The question to ask: “Which section of this document would the company least want a journalist to read, and why?”

Section to question: how to read a card in 30 minutes

Card section	The one question to ask	What a red flag looks like
Model details	Which exact version is my kid using?	Card describes a version the product does not run
Intended use	Is education or minors listed as in scope?	Silence about minors on a product marketed to schools
Factors	Was performance broken out by language and age?	A single global average
Metrics	Benchmark score at what cost and with what prompting?	A number with no benchmark named
Evaluation data	Does the test data look like my kid’s homework?	Only English, only adults, only synthetic
Training data	What is disclosed, and what is withheld?	One vague paragraph
Safeguards / jailbreaks	What percentage of jailbreak attempts succeed?	No number at all
Alignment / honesty	Did any metric get worse than the last version?	Only improvements reported
Agentic safety	What happens when it has tools and a browser?	Not addressed on an agentic product
Caveats	What do the makers say not to trust it for?	Section absent or two sentences

Print that table. It works on any model card from any company, and it works better than reading the document front to back.

What to actually do at home

Read the card for the tool your kid actually uses

Not the most famous model. The one in their homework app or their district’s rollout. Utah’s statewide Gemini partnership covers roughly 680,000 students and 28,000 educators at no cost for 2026-27; families in that state should know which model card applies. Most parents have never asked.

Ask your school which model card they reviewed

This single question changes conversations. The US Department of Education’s Dear Colleague Letter of August 20, 2026 told districts to evaluate AI by measurable learning outcomes rather than screen time. A district that cannot name the model card it reviewed has not done that evaluation. Ask politely, in writing, once.

Use disclosed regressions as a trust signal, not a scare

A company that publishes “this got worse” is behaving better than one that publishes only wins. Teach your teen to read a disclosed weakness as evidence of good process. This is counterintuitive and it is correct, and it transfers to how they should read scientific papers and their own lab reports.

Pair the card with the marketing page

Open both side by side. The launch blog post says “most capable model.” The system card says jailbreak success is 4.5 percent and honesty under pressure declined. Neither is lying. They are answering different questions for different audiences. Recognizing that is media literacy at a level most adults have not reached. See our pieces on Claude Fable 5.1 and Terminal-Bench-Science and how safety classifiers decline requests.

What not to do

Do not try to read 200 pages. You will stop, feel bad, and conclude the document is useless. Use the table, pick three sections, and be done in half an hour. Depth is not the goal; the habit of looking is.

What to Watch For Over the Next 3 Months

• Week 4: Your teen has located the caveats section of one real model card and can say what it admits.
• Month 2 red flags: A tool your kid uses at school has no published card, or the card describes a different version than the one deployed.
• Month 3 self-check: When a new model launches, does anyone in your house ask what the system card says? If the reflex has formed, this worked.

Frequently Asked Questions

What is the difference between a model card and a system card?

A model card, as proposed in 2019, documents a model: intended use, metrics, evaluation, limitations. A system card documents a deployed system, which includes the model plus safeguards, monitoring, tool access, and a risk determination. Frontier labs now publish system cards because the safeguards are as important as the weights.

Are companies required to publish these?

Not universally. The EU AI Act and the EU Code of Practice on Transparency of AI-Generated Content push in that direction, and many US agencies expect documentation for procurement. Most frontier labs publish voluntarily because customers and regulators ask. The practice is normalizing faster than the law.

Should I trust the numbers in a system card?

Treat them as the company’s best self-report, produced with genuine internal rigor and real incentive to look good. The presence of disclosed regressions is a positive signal. Independent verification remains scarce, which is a genuine gap in the field and worth saying out loud.

My kid’s school uses an AI tool with no model card. Is that bad?

It is a reason to ask questions, not to panic. Many education tools are built on top of a frontier model, so the underlying model’s card may exist even if the app’s does not. Ask the vendor which model they use and which version. If they cannot answer, that is the finding.

Where do I find these documents?

On the maker’s site, usually under a research, safety, or documentation section. Anthropic publishes system cards alongside model announcements; OpenAI publishes system cards for major releases; Google publishes model cards for Gemma and Gemini. Search the model name plus “system card.”

Is 4.5 percent jailbreak success high or low?

Low compared to models of a few years ago, and not zero. In a school with 1,000 students, a determined minority will find that 4.5 percent. The useful conclusion is that safeguards reduce casual misuse substantially and do not eliminate deliberate misuse, which is exactly how to think about every safety layer.

---

About the author

Ricky Flores is the founder of HiWave Makers and an electrical engineer with 15+ years of experience building consumer technology at Apple, Samsung, and Texas Instruments. He writes about how kids learn to build, think, and create in a tech-saturated world. Read more at hiwavemakers.com.

---

Sources

1. Mitchell, M., Wu, S., Zaldivar, A., Barnes, P., Vasserman, L., Hutchinson, B., Spitzer, E., Raji, I. D., & Gebru, T. (2019). “Model Cards for Model Reporting.” FAT ‘19*. https://arxiv.org/abs/1810.03993
2. Anthropic. (2026, September 1). “Introducing Claude Fable 5.1 and Claude Mythos 5.1.” https://www.anthropic.com/claude-fable-and-mythos-5-1
3. Anthropic. (2026, September 1). “System Card: Claude Fable 5.1 & Claude Mythos 5.1.” https://www-cdn.anthropic.com/0339e6a7c5c7b87f5c07798616dc32c215d14235/Claude%20Fable%205.1%20&%20Claude%20Mythos%205.1%20System%20Card.pdf
4. Don’t Worry About the Vase. (2026, September 4). “Claude Fable 5.1 and Mythos 5.1: The System Card.” https://thezvi.wordpress.com/2026/09/04/claude-fable-5-1-and-mythos-5-1-the-system-card/
5. MarkTechPost. (2026, September 1). “Anthropic releases Claude Fable 5.1 and Claude Mythos 5.1.” https://www.marktechpost.com/2026/09/01/anthropic-releases-claude-fable-5-1-and-claude-mythos-5-1-52-6-on-terminal-bench-science-and-75-cheaper-cache-reads/
6. US Department of Education. (2026, August 20). Dear Colleague Letter on evaluating AI by measurable learning outcomes. https://www.ed.gov/
7. Terminal-Bench. “Terminal-Bench-Science 0.1.” https://www.tbench.ai/news/terminal-bench-science-0-1