Test-Time Compute Explained: Why AI 'Thinks Longer' on Math

Test-time compute is the amount of computation a model spends after your question arrives and before it answers. Not training. Not model size. Just thinking time on this particular problem. It’s the reason multiple AI systems scored a perfect 42/42 at the International Mathematical Olympiad in Shanghai in July 2026, a competition where only 7 of 666 human contestants achieved the same. And it’s the reason your kid’s homework helper sometimes pauses for 90 seconds and comes back right, when the fast version came back confidently wrong.

Key Takeaways

• Test-time compute is effort spent at answer time: generating multiple candidate solutions, checking them, revising, and picking the best.
• Snell et al. (2024) found that compute-optimal allocation at test time beat standard best-of-N sampling by more than 4x in efficiency, and on some problems a small model with extra thinking time outperformed a model 14 times larger.
• At IMO 2026 (Shanghai, July 10–21), multiple AI systems were graded at a perfect 42/42, including Huawei’s Celia and RedNote’s dots-note-3.0. Among humans, 55 golds and 7 perfect scores out of 666 contestants.
• More thinking time is not free: it costs money, it costs latency, and it hits diminishing returns. The right amount depends on the problem’s difficulty.
• The parenting lesson is the transferable one: the fast answer and the careful answer are different products, and a kid should learn to ask for the right one.

What test-time compute means, mechanically

When you send a question to a modern AI model, three separate budgets have already been spent or are about to be. Training compute was spent months ago, teaching the model patterns from text. Model size is fixed at that point: a certain number of parameters, a certain cost per token. Test-time compute is the only budget still open.

Here’s what the model can actually do with it:

Generate more than one answer. Sample the same question ten times and you get ten attempts, because the sampling process is probabilistic. This is called best-of-N.

Check its own work. A verifier, often a separate reward model trained to score partial solutions, evaluates candidates. On math problems where steps can be graded, this works well.

Revise. The model reads its own draft, spots an error, and rewrites. Sequential revision is different from parallel sampling, and the research shows they suit different problem types.

Search. Instead of writing one chain of reasoning straight through, explore a tree of possible next steps and evaluate branches.

The foundational paper here is Snell, Lee, Xu, and Kumar, “Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters”, posted August 2024. Two findings matter for a parent. First, how you spend the budget matters enormously: a compute-optimal strategy beat naive best-of-N sampling by “more than 4x” in efficiency. Second, and more surprising, on some problem categories a smaller base model given extra test-time computation “outperform[ed] a 14x larger model” at matched total compute.

Third finding, and the practical one: effectiveness depends on prompt difficulty. Easy problems get nothing from extra thinking. Hard ones get a lot. So the optimal strategy is adaptive, which is exactly what “adaptive thinking” in current model documentation refers to.

What happened at the 2026 Math Olympiad

The International Mathematical Olympiad ran July 10–21, 2026, in Shanghai, with 666 contestants from 117 countries. Fifty-five gold medals were awarded, at a threshold of 29 points, and 7 contestants achieved the maximum 42.

On the AI side, the numbers are unambiguous and the accounting is messy. Huawei’s Celia and RedNote’s (Xiaohongshu’s) dots-note-3.0 were graded at 42/42. South China Morning Post reported on July 22 that dots-note-3.0 hit a perfect score, noting that the previous year Google DeepMind and OpenAI had managed 35/42. Several other labs, including Anthropic and OpenAI, reported 42/42 under self-administered conditions rather than official grading, as catalogued in one detailed accounting of the results. Sources differ on the exact count of systems at 42, so “multiple” is the honest word.

The mechanism behind the jump from 35 to 42 in one year is mostly test-time compute. An IMO problem takes a strong human contestant roughly 90 minutes. A model that answers in three seconds is not doing the same thing a model that spends 40 minutes generating, verifying, and discarding approaches is doing. The second one is spending test-time compute.

Two honest caveats belong here. Olympiad problems have verifiable answers, which is precisely the setting where checking your own work pays off, a point made carefully in the survey “From Solvers to Research”. And a maxed-out score tells you less than a partial one, because once several systems hit the ceiling the number stops distinguishing them. Our piece on what a maxed-out benchmark can’t tell you works through that problem, and how AI borrowed number theory to crack a geometry conjecture covers the research-math side.

Why this changes what “smart” means for a model

For most of the last decade, the answer to “make the model better” was “make the model bigger.” Test-time compute broke that monopoly. A cheaper, smaller model that thinks longer can now match a bigger one on some tasks, and that has three consequences a family will actually notice.

Price and time became a dial you control. Most AI products now expose some version of a thinking-effort setting, and the published per-token pricing shows what that dial costs, since deliberation is billed as output. Choosing it is a real decision, not a cosmetic one.

Smaller models got more useful. Which matters for schools, for on-device AI on a phone (Google’s August 2026 roundup describes Gemini Nano running on the Pixel 11’s Tensor G6), and for anyone on a budget. Our explainer on why AI models come in different sizes covers the tiers.

The failure modes changed. A model rushing gives you a confident wrong answer. A model thinking gives you a slower, usually better answer, and occasionally an elaborate wrong answer with 400 words of reasoning attached. The second kind is harder for a kid to catch, because it looks like work.

That last point is the one to actually worry about. Reasoning text is persuasive. A 13-year-old checking a math answer against a long chain of steps often assumes the length is evidence. It isn’t.

How to Teach Your Kid About Test-Time Compute

The concept is “how long did you think about it,” which children understand intuitively. The trick is making the trade-off visible.

Ages 5–8: Fast hand, slow hand

Ask a mental math question (8 + 7). Have your kid answer as fast as possible. Then ask a harder one (14 + 19) and require them to count on their fingers or use paper, out loud, slowly. Ask: “Which one needed more thinking?” Then say: “Computers are the same. Some questions they answer instantly. For hard ones, they have to think longer, and it costs their owner more money when they do.” Try a question where the fast answer is wrong (17 + 8, answered in under a second) and let them catch it themselves.

Ages 9–12: The three-tries experiment

Give your kid a multi-step word problem with a checkable answer. Round one: 30 seconds, one attempt, no checking. Round two: three minutes, and they must solve it three different ways and compare. Record both answers in a table. Then explain that round two was best-of-N with verification, and that this is roughly what changed between AI scoring 35 out of 42 and 42 out of 42 at the Olympiad. Do it three times with different problems. Count how often the fast answer and the careful answer disagree.

Ages 13+: Measure it yourself

Have your teen pick five problems of increasing difficulty (an arithmetic problem, an algebra problem, a competition-style problem, a proof, an open-ended question). For each, run it through an AI tool twice: once asking for an immediate answer, once asking it to work carefully and check its reasoning. Record the answer, the time, and whether it was right. Then plot difficulty against the gap between the two. They should find what Snell et al. found: no benefit on easy items, a real gap on hard ones. That’s a genuine replication of a published result, done at a kitchen table.

The question to ask: “When would you want the fast answer even if it might be wrong, and when would you pay for the slow one?”

Thinking time vs. accuracy: how the trade-off actually behaves

Thinking budget	What the model does	Where it helps	Where it doesn’t	Cost side effect
Minimal (one pass)	Generates one answer straight through	Simple lookups, formatting, easy arithmetic	Multi-step reasoning, proofs	Cheapest and fastest
Low (a few samples)	Samples several answers, picks the most common	Moderate word problems	Anything needing a novel approach	Small multiple of base cost
Medium (sample + verify)	Generates candidates, scores them with a verifier, picks the best	Math with checkable answers; code that can be run	Open-ended writing with no ground truth	Noticeably slower; more tokens billed
High (search + revise)	Explores branches, revises drafts, discards dead ends	Olympiad problems, hard proofs, long agent tasks	Easy problems (pure waste)	Can be 10x+ the tokens; minutes of latency
Beyond optimal	Keeps going past the useful point	Nothing	Everything	Diminishing returns; you pay for no gain

The row that surprises people is the last one. More thinking is not monotonically better. Snell et al.’s core argument was about optimal allocation, not maximum allocation, and the optimum depends on how hard the question is.

What to do at home

Name the two modes out loud

“Do you want the quick answer or the careful answer?” is a question worth asking your kid about their own homework, not just about the AI. The habit of choosing a mode deliberately is the transferable skill, and it applies to writing a paragraph as much as to solving for x.

Treat long reasoning as a claim, not proof

When an AI produces a wall of steps, have your kid check one step at random rather than reading the whole thing. Middle steps are where errors hide, and spot-checking is a real technique. Our guide on building a weekly ritual for verifying AI math sets up the routine.

Use the thinking setting on purpose

If your kid’s tool has an effort or thinking control, teach them to turn it up for math and proofs and down for summaries and formatting. This is the practical payoff of understanding test-time compute, and it saves real money on paid plans.

Time the model, then time yourself

Have your kid solve a problem while the AI works on the same one. Compare both answers and both durations. Kids who have raced a model on a problem they can actually do develop a much more accurate sense of where it’s strong.

What not to do

Don’t teach that thinking longer always means being right. That’s the exact error the research disproves, and it primes a kid to trust elaborate wrong answers. And don’t let the Olympiad headline become “AI is better at math than humans.” Competition math with verifiable answers is one narrow slice, and 7 human contestants also scored 42/42 while writing by hand in a room with no verifier.

What to Watch For Over the Next 3 Months

• Week 4: Your kid can explain in one sentence that test-time compute is thinking after the question, not learning before it, and can name one task where it helps.
• Month 2 red flags: They trust long reasoning chains without checking a step. Or they’ve concluded that a slow answer is automatically better than a fast one.
• Month 3 self-check: Ask them to predict, for three problems of different difficulty, whether extra thinking time will change the answer. Getting two right with a reason is a pass.

Frequently Asked Questions

Is test-time compute the same as the model being smarter?

No. It’s effort at answer time, separate from training and model size. The striking result from Snell et al. (2024) is that on some problems a smaller model with a well-spent thinking budget outperformed a model 14 times larger. Capability and effort are different levers.

Why does my AI tool sometimes take 90 seconds?

Because it’s spending test-time compute: sampling multiple approaches, checking them, or revising. On hard problems that usually buys accuracy. On easy ones it buys nothing, which is why newer models try to allocate the budget adaptively.

Did AI really beat humans at the Math Olympiad?

Multiple AI systems were graded at a perfect 42/42 at IMO 2026, including Huawei’s Celia and RedNote’s dots-note-3.0. Seven of 666 human contestants also scored 42/42, under strict time limits and by hand. The AI results are real, and “beat humans” flattens meaningful differences in conditions.

Should I pay for the higher-thinking tier for my kid’s homework help?

It depends on the task. Math, proofs, and multi-step problems benefit measurably. Summarizing an article or fixing grammar does not. Paying for maximum thinking on every request is the most common way families waste money on AI tools.

Does thinking longer reduce hallucinations?

Sometimes, and not reliably. Verification helps most when there’s a ground truth to check against, like a math answer or code that runs. For factual claims with no built-in check, a model can spend a long time producing a confident, well-structured, wrong answer.

---

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. Snell, C., Lee, J., Xu, K., & Kumar, A. (2024). “Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters.” arXiv:2408.03314. https://arxiv.org/abs/2408.03314
2. International Mathematical Olympiad. (2026). “IMO 2026, Shanghai, People’s Republic of China.” https://www.imo-official.org/editions/2026/
3. South China Morning Post. (2026, July 22). “World’s first AI model to earn perfect score in maths olympiad comes from China’s RedNote.” https://www.scmp.com/tech/article/3361482/worlds-first-ai-model-earn-perfect-score-maths-olympiad-comes-chinas-rednote
4. Digital Applied. (2026, July 23). “IMO 2026 Perfect Scores and AI Benchmark Saturation.” https://www.digitalapplied.com/blog/imo-2026-perfect-scores-ai-benchmark-saturation
5. TechXplore. (2026, July). “AI systems match humans’ score in prestigious math contest.” https://techxplore.com/news/2026-07-ai-humans-score-math-contest.html
6. Anthropic. (2026). “Pricing.” Claude Platform Documentation. https://platform.claude.com/docs/en/about-claude/pricing
7. Science News. (2026, June 8). “AI guardrails and the Erdős math problem.” https://www.sciencenews.org/article/ai-guardrails-erdos-math-problem