AI Marketing Numbers: Teaching Kids to Read the Claim

“54% more token efficient.” Sam Altman said it about GPT-5.6 Sol in July 2026, tech press repeated it, and almost nobody asked the obvious follow-up: more efficient than what, measured how, on which task? The answer turns out to be specific and mostly defensible, which is exactly why it makes such a good lesson. Teaching kids to read AI marketing numbers works best on a claim that is partly true, because the skill is not detecting lies. It is finding the boundary of what a number covers.

Your kid already reads numbers like this every day, in game ads, supplement labels, and school test scores.

Key Takeaways

• The precise claim, from OpenAI’s developer account, is that GPT-5.6 Sol at max reasoning reached 80 on the Artificial Analysis Coding Agent Index “while using 54% fewer output tokens and completing tasks in 57% less time than the next-highest-scoring model.”
• That is a comparison against one competing model on one benchmark in one harness, not a general statement about efficiency.
• Independent measurement from Artificial Analysis showed Sol using about 15,000 tokens per Intelligence Index task against GPT-5.5’s 16,000, roughly a 6% improvement on a different measure.
• Five questions, compared with what, measured by whom, on which task, how many runs, and does it hold at my settings, catch most inflated claims.
• Kids need this skill: Breakstone et al. (2021) found 96% of 3,446 high school students failed to notice a climate-information site’s fossil-fuel ties.

AI marketing numbers kids hear: the claim unpacked word by word

Start with the exact wording, because the vaguer version is the one that travels. Reporting on the GPT-5.6 launch on July 9, 2026 carried Altman’s line that Sol is 54% more token efficient for AI coding tasks. OpenAI’s own developer account was more precise: on the Artificial Analysis Coding Agent Index, “Sol with max reasoning sets a new high of 80 while using 54% fewer output tokens and completing tasks in 57% less time than the next-highest-scoring model.”

Now the unpacking, one clause at a time.

“54% fewer output tokens” is a specific, measurable quantity. Output tokens are what the model writes, including its reasoning, and they are the expensive half of an API bill, as our tokens and per-million pricing explainer lays out. Fewer output tokens for the same score is a genuine engineering achievement, not a vibe.

“than the next-highest-scoring model” is the comparison, and it is doing a lot of work. The next-highest model on that index was Claude Fable 5. So the claim is about one competitor, not an average of the field.

“on the Artificial Analysis Coding Agent Index” is the scope. Coding agent tasks. Not writing, not math tutoring, not reading comprehension. Fifty-four percent fewer tokens at coding says nothing about how many tokens the model spends explaining fractions.

“with max reasoning” is the setting. The number applies at the highest reasoning effort, which is the most expensive configuration. A family running the same model at default settings gets different numbers.

“sets a new high of 80” is the headline result, and the honest version is that 80 was 2.8 points above Fable 5. Better, clearly. Not a different league.

So is the claim true? Mostly yes, as stated, in context. Is the version your kid will hear, “the new AI is 54% more efficient,” true? Not really, because it drops all four qualifications.

What an independent measurement showed

This is the part that makes the lesson land. Artificial Analysis published its own evaluation on the same day, and on a different metric, tokens per Intelligence Index task, Sol used about 15,000 tokens against GPT-5.5’s 16,000. That is roughly a 6% improvement.

Both numbers are real. They measure different things: OpenAI’s compares Sol to a competitor on a coding-agent index; the independent one compares Sol to its own predecessor on a general intelligence index. A kid who can hold both in their head, without concluding “so it was all a lie,” has acquired the actual skill.

The same pattern shows up everywhere in AI marketing. Anthropic reported 52.6% for Claude Fable 5.1 on Terminal-Bench-Science in September 2026, while the public leaderboard listed 21.4% and 30.0% for closely related configurations, because harness and effort settings differ. VentureBeat’s coverage of that launch put it plainly: vendor-reported results are “not independent proof of superiority,” and “production safeguards can affect scores.” Notice that the honest caveat came from the trade press, not the press release. Our walkthrough of that benchmark shows why both numbers can be accurate.

How to Teach Your Kid About AI Marketing Numbers

Ages 5–8: The “more than what?” game

Every time a package or an ad says “more,” “faster,” or “better,” ask your child: more than what? Cereal boxes are perfect training material. Do it for a week and a five-year-old will start asking it unprompted, which is the whole goal. Numbers need a partner to compare against, and most ads hide the partner.

Ages 9–12: Make a misleading poster on purpose

Give your kid a true fact about themselves and have them write three headlines from it: one honest, one that exaggerates without lying, and one that is false. Example fact: they ran a mile 20 seconds faster than last month. Honest: “improved by 20 seconds.” Exaggerated but not false: “dramatically faster runner.” False: “fastest in the school.” Then have them explain why the middle one is the dangerous one. Kids who have built the trick recognize it afterward.

Ages 13+: Audit a real claim

Hand your teen the 54% line and the five questions below. Have them write a one-paragraph verdict: what the claim supports, what it does not, and what they would need to check it. Then show them the Artificial Analysis 15,000-versus-16,000 figure and ask whether it contradicts the 54% claim. The correct answer is no, and working out why is the entire lesson in scope and measurement.

The question to ask: “This number is true. What would someone wrongly believe if they only heard the short version?”

The five questions, applied to real 2026 claims

Claim	Question to ask	What the answer reveals
”54% more token efficient” (GPT-5.6 Sol)	Compared with what, on which task, at which setting?	Against one competitor, on a coding-agent index, at max reasoning
”Close to the frontier at half the price” (Claude Opus 5)	Half the price of what, and close on which benchmark?	$5/$25 versus Fable 5’s $10/$50; within 0.5% on CursorBench at max effort
”52.6% on Terminal-Bench-Science” (Fable 5.1)	Who ran the test, in which harness?	Anthropic’s own harness; the public leaderboard lists lower figures for related setups
”Half the price of its predecessor” (Gemini 3.7 Flash)	For how long?	Introductory pricing that expires December 31, 2026, then doubles
”Most aligned model to date” (Claude Opus 5)	Aligned by which measurement?	A score of 2.3 on Anthropic’s own automated behavioral audit
”AI tutor doubled learning gains” (Kestin et al. 2025)	Who was the control group?	An active-learning classroom, which is itself best practice, making the result stronger than it sounds

Note the last row, because it cuts the other way. Sometimes the full context makes a claim more impressive, not less. In Kestin et al. (2025), published in Scientific Reports, 194 Harvard physics students using a carefully designed AI tutor showed median learning gains more than double the control group, and the control group was an active-learning classroom rather than a lecture. Asking “compared with what” is not a way to debunk. It is a way to find out.

Why this skill is worth the effort

The evidence on kids’ evaluation of online claims is not encouraging. Breakstone, Smith, Wineburg et al. (2021), published in Educational Researcher, assessed 3,446 U.S. high school students with live internet access. Ninety-six percent never discovered that a site claiming to publish factual climate reports had ties to the fossil fuel industry. Two-thirds could not distinguish news stories from ads on a homepage. More than half judged an anonymously posted video shot in Russia to be strong evidence of U.S. voter fraud.

Those students had the internet in front of them. The missing piece was not access; it was the habit of asking who is telling me this and how would they know.

AI claims are unusually good practice material for two reasons. They are quantitative, so the qualifications are findable rather than a matter of taste. And they are checkable, because independent evaluators like Artificial Analysis and public leaderboards publish their own numbers. A teen can actually resolve the question, which almost never happens with political claims.

Meanwhile the OECD’s PISA 2025 results, released September 8, 2026, found that 46% of students in OECD countries use AI chatbots at least weekly for learning, while only 27% of U.S. teens in Common Sense Media’s August 2026 survey said a teacher had ever discussed safe AI use with them. Kids are inside the marketing whether or not anyone teaches them to read it.

What to Watch For Over the Next 3 Months

• Week 4: Catch one AI claim in the wild together, from an app store listing or a YouTube ad, and run the five questions on it.
• Month 2 red flags: Your kid repeating a percentage with no comparison group; a cited finding with no named researcher; confusing a vendor’s number with an independent test.
• Month 3 self-check: New models will ship with new records. Ask your kid to predict what the qualification will be before you read the announcement together. Getting good at that prediction is the sign it stuck.

Frequently Asked Questions

Is the “54% more token efficient” claim false?

No. The precise version, that GPT-5.6 Sol at max reasoning scored 80 on the Artificial Analysis Coding Agent Index while using 54% fewer output tokens than the next-highest-scoring model, appears accurate. What is misleading is the shortened version, which drops the comparison model, the benchmark, and the reasoning setting.

Why did an independent test show only about 6%?

Because it measured something else. Artificial Analysis compared tokens per Intelligence Index task between Sol (about 15,000) and GPT-5.5 (about 16,000). OpenAI’s figure compared Sol to a competitor on a coding-agent index. Different comparison, different task, different number.

What are the five questions to ask about any AI claim?

Compared with what? Measured by whom? On which task or benchmark? Across how many runs? And does it hold at the settings I would actually use? Those five catch most of the gap between a claim and what people hear.

At what age can kids learn this?

Five-year-olds can learn “more than what?” from a cereal box. Nine-to-twelves can build a misleading headline on purpose. Teenagers can audit a real claim against an independent source. The skill scales; only the material changes.

Are AI companies unusually dishonest about numbers?

Not especially. Their claims are often more checkable than the industry average, because benchmarks are public and independent evaluators publish. The problem is that the qualifications are technical, so press coverage strips them, and the stripped version is what circulates.

How do I check a benchmark claim myself?

Look for the benchmark’s own site and leaderboard, note the harness and effort setting listed next to each score, and compare the vendor’s figure to the leaderboard’s. If a vendor’s number is much higher, the configuration usually explains it, and the configuration is part of the claim.

---

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. TechCrunch. (2026, July 9). “OpenAI launches its new family of models with GPT-5.6.” https://techcrunch.com/2026/07/09/openai-launches-its-new-family-of-models-with-gpt-5-6/
2. Artificial Analysis. (2026, July 9). “GPT-5.6 benchmarks across Intelligence, Speed and Cost.” https://artificialanalysis.ai/articles/gpt-5-6-has-landed
3. OpenAI. (2026, July 9). “GPT-5.6: Frontier intelligence that scales with your ambition.” https://openai.com/index/gpt-5-6/
4. Anthropic. (2026, July 24). “Introducing Claude Opus 5.” https://www.anthropic.com/news/claude-opus-5
5. VentureBeat. (2026, September 1). “Anthropic’s Claude Fable 5.1 and Mythos 5.1 arrive with a 75% cost reduction for Fable cache reads.” https://venturebeat.com/technology/anthropics-claude-fable-5-1-and-mythos-5-1-arrive-with-a-75-cost-reduction-for-fable-cache-reads
6. Breakstone, J., Smith, M., Wineburg, S., et al. (2021). “Students’ Civic Online Reasoning: A National Portrait.” Educational Researcher, 50(8), 505–515. https://journals.sagepub.com/doi/10.3102/0013189X211017495
7. Kestin, G., Miller, K., Klales, A., et al. (2025). “AI tutoring outperforms in-class active learning: an RCT introducing a novel research-based design in an authentic educational setting.” Scientific Reports. https://www.nature.com/articles/s41598-025-97652-6
8. OECD. (2026, September 8). “PISA 2025: Students’ reading and mathematics performance declined sharply across the OECD.” https://www.oecd.org/en/about/news/press-releases/2026/09/pisa-2025-students-reading-and-mathematics-performance-declined-sharply-across-the-oecd.html
9. Common Sense Media. (2026, August 18). “Teens in the AI Era: Schoolwork and the Skills That Matter.” https://www.commonsensemedia.org/research/teens-in-the-ai-era-schoolwork-and-skills-that-matter