Reasoning Models in Schools: Chain-of-Thought, Adaptive Thinking, and What Changes for Marking and STEM

Period seven on Wednesday, a Form 5 physics teacher runs a first pass on lab reports in the staffroom. The model works through the rubric criterion by criterion, even quoting the student's own phrase about "controlled variables" — far richer feedback than last term. She still deducts marks for data handling: the model treats a miscopied resistance unit as reasonable and smooths the error into fluent commentary. She mutters at the screen: "It writes report feedback now. It doesn't run the experiment."

Reasoning models turned AI from a contestant who blurts out answers into one who drafts on paper before handing in the script. That is a real gain for marking and multi-step explanation. A neat draft with wrong facts is still wrong, and the teacher checkpoint stays.

Three generations of thinking

First-generation language models behaved like an oral exam: answer immediately, think while speaking, and lose the thread on anything multi-step. Teachers then added "think step by step" to prompts so the model wrote intermediate steps before the conclusion — chain-of-thought in classroom terms, giving the model a visible working sheet rather than a new theory.

Through 2024 and 2025 the working sheet moved backstage. OpenAI's o1, o3 and o4-mini reasoned internally before output; on 20 January 2025 DeepSeek-R1 published an open-weight equivalent; on 24 February 2025 Anthropic's Claude 3.7 Sonnet exposed extended thinking as a switch. Users mostly saw only the answer, or a brief summary.

2026 brought adaptive thinking: the model judges whether a question warrants effort and for how long. Anthropic calls it adaptive thinking; Google defaults to dynamic thinking. Vendor documentation is consistent: the effort parameter is behavioural guidance, not a hard ceiling — even at the lowest setting, a hard enough problem can still trigger deep thinking.

The four vendors' effort dials

The following describes design as published in September 2026 documentation, not a performance ranking.

OpenAI uses reasoning.effort from low through max (GPT-6 Astra does not support none); GPT-5.6 defaults to medium, with independent standard and pro reasoning modes — the finest-grained control.

Anthropic pairs thinking with effort at low, medium, high, xhigh and max; adaptive thinking is always on for Claude Fable 5.1 with effort defaulting to high, which suits judgement-heavy feedback.

Google uses thinking_level (low / medium / high) with dynamic thinking on by default; documentation advises lowering thinking_level to control cost rather than shrinking max_output_tokens.

xAI offers reasoning_effort at low, medium, high and xhigh, defaulting to high — fewer levels, simpler configuration.

Effort is not a cost guarantee

Anthropic's documentation is explicit: at lower effort the model still thinks when a problem is hard enough, and the real hard limit is max_tokens. A school wiring the API directly without routing can burn a month's budget on one batch marking run.

The thinking-token bill

Thinking tokens charge at output rates. Flagship output is typically five times input. Moving from medium to xhigh can double output tokens — in 2026 the easiest cost to lose control of is not "more chats" but "each chat thinking too long".

Three staffroom shifts

Marking feedback

2024-era AI feedback often read like generic boilerplate — "content adequate, structure could improve" — without genuinely comparing against the rubric. Reasoning models work through criteria first, locate matching sentences in the student's text, and score with markedly better consistency across repeated runs; feedback starts quoting the student's own words. That is what moved AI marking from demo to daily use.

Marks still require teacher review before release. The model does not know what you emphasised last week, or a particular SEN arrangement.

Maths and science lessons

Multi-step calculation reliability has improved, and intermediate steps can be formatted straight onto a worksheet. Asking for the same problem at three difficulty levels helps differentiated teaching.

The risk sits here too: a complete, orderly derivation can still reach a wrong conclusion and is now harder to spot. The guidance is one sentence: any worked solution going to students must be checked step by step, not only at the final answer.

Scientific explanation

Causal chains and "what if we change the assumption" reasoning are where reasoning models help science teaching most. Keep two things apart: reasoning improves getting from the known to the unknown; it does not make the model know more. It still misremembers experimental values, textbook editions and Hong Kong curriculum unit sequencing in the same confident tone.

Reasoning makes wrong answers look more like right ones — more dangerous than messy errors, because a teacher might glance and approve.

Four holes reasoning cannot fill

Knowledge cutoffs differ by vendor: 30 April 2026 for GPT-6 Astra, 1 February 2026 for Grok 4.6, January 2026 for Claude Sonnet 5. No amount of thinking reaches past the cutoff; where current facts matter, answers should come from the school knowledge base.

Local knowledge gaps — Hong Kong curriculum codes, EDB circular numbers, full school names — are thin in any training corpus; reasoning only organises the wrong content better.

Classroom latency: while the model thinks, the student waits. Fifteen seconds is enough to lose a Form 3 class; student-facing chat should not run at high effort across the board.

Audit visibility: thinking content is usually invisible or summarised. Audit records should preserve input, output and model settings rather than assuming full internal reasoning can be reconstructed later.

Platform routing and rewritten prompts

Schools need not tune parameters themselves, but should know what the platform does: classification, tagging, reformatting and notice drafts at the lowest effort; planning, question generation and summarising at low to medium; composition feedback and rubric alignment at medium to high; cross-document policy cross-reference and multi-step STEM explanation only where worth paying for; live student chat at low to medium because latency beats depth.

Prompts need rewriting too. Instead of "think step by step", state the task, output fields, language and usage, and flag uncertainty rather than filling gaps. OpenAI's reasoning best practices note that teaching the model how to think may not help reasoning models and can hinder them; see the prompt engineering course.

A practical opener: "You are a Hong Kong secondary physics teacher writing for Form 5 students. For the lab report below, write feedback against each rubric criterion; output: criterion number, quoted student sentence, one suggestion each; constraints: Traditional Chinese with Hong Kong usage, mark data or units as 'needs verification' rather than correcting them yourself." That last line drops review cost from redoing the work to scanning line by line.

Common misreadings

• The claim: Reasoning models stop getting things wrong. What actually happens: Reasoning reduces skipped steps and carelessness, not factual error. A tidy derivation can still reach a wrong conclusion and is harder to catch.

• The claim: The same question answering fast one moment and slow the next means the system is unstable. What actually happens: 2026 models decide how long to think. Effort is guidance, not a ceiling; hard problems can still trigger deep thinking — expected behaviour.

• The claim: Turning reasoning effort all the way up always gives the best quality and the school can afford it. What actually happens: Output tokens (including thinking) bill at output rates, typically five times input on flagships. Without routing, the cost curve is non-linear.

Three things to try on Monday

1. Take one assignment with an existing rubric, run an AI first pass, and check whether feedback quotes the student's own sentences; if you only get boilerplate, verify the rubric is uploaded to the platform.
2. Pick one multi-step maths or science problem, ask for working where "every step is independently checkable", and verify step by step rather than only the final answer.
3. Confirm with the ICT coordinator that student chat runs at lower reasoning effort and batch marking is routed by task, so classroom waits and runaway bills are avoided.