Skip to main content
HSC 2026 final syllabus year published Aug 2026

Mathematics Standard 2 question styles to look out for

The question styles to have ready — each linked to its evidence and to a question in the practice paper.

Built by a six-model AI panel and backtested against the hidden 2025 papers — how we did it.

About the 2026 exam & how this page was built

Built from a six-model AI analysis of every HSC Maths Standard 2 paper, marking guideline and marking-centre feedback report since 2019 — the same method backtested against the real 2025 papers in chemistry and maths ext 1 before publishing. 2026 is the FINAL year of the current syllabus, and the format has been identical since 2019: 15 multiple-choice marks, then 85 marks of short questions (Q16–40), 2.5 hours. Final-year papers historically favour well-trodden formats — which makes style preparation unusually valuable this year.

The near-certainties

Section II · 3–6 marks 6 of 6 models expect this

Regression with the interpretation marks (3–6 marks)

The panel's only unanimous cluster.

Try it → Q25 Why we expect this
Section II · 3–5 marks 5 of 6 models expect this

The quadratic model with a maximum (3–5 marks)

A supplied parabola modelling revenue, enclosure area, braking distance or a projectile: locate the turning point from the axis of symmetry (or given intercepts), evaluate the maximum,...

Try it → Q22 Why we expect this
Section II · 3–5 marks 5 of 6 models expect this

Networks, both flavours

A distance/cable table between five to eight towns is first drawn as a weighted network, the MST is drawn on a supplied blank diagram and its total length stated; the follow-up perturbs...

Try it → Q19 Why we expect this
On this page
  1. 1. The quadratic model with a maximum (3–5 marks)
  2. 2. Networks, both flavours
  3. 3. Bearings navigation returns (4–5 marks)
  4. 4. The interest-factor table (annuities)
  5. 5. Regression with the interpretation marks (3–6 marks)
  6. Where to spend study time
  7. Practice paper
  8. Check our working

The near-certainties (the strongest cross-model agreement)

1. The quadratic model with a maximum (3–5 marks)

5 of 6 models expect this — consensus probability 0.70

The panel's strongest question-level call (five models, 0.70): revenue, enclosure area, braking distance or a projectile, with the parabola supplied. Find the turning point from the axis of symmetry, evaluate the maximum, then either find the second input giving the same output or state the sensible domain. The traps flagged in 2021, 2024 and 2025: reading the vertex where an intercept is wanted, and ignoring that a width can't be negative.

Practise

one quadratic-model question per week, always finishing with a sentence about the practical domain.

Try it → Q22 5 of 6 models expect this — see the evidence

2. Networks, both flavours

5 of 6 models expect this — consensus probability 0.69

Minimum spanning tree from a distance table (five models, 0.69): draw the network from the table first, then the MST, then survive the perturbation — a road closure forcing a new shortest path, or "is a second tree of equal length possible?". The designed trap every time it runs (2019, 2023, 2024): the shortest path does not have to lie inside the MST. Critical path (four models, 0.69): forward/backward scan, minimum completion time, then $\text{float} = \text{LST} - \text{EST}$ used to justify whether a changed duration alters the critical path. One caution: this is the panel's only contested topic — two models rated it rested — so treat it as very likely, not certain.

Practise

networks with a written justification sentence, not just numbers.

Try it → Q19 5 of 6 models expect this — see the evidence

3. Bearings navigation returns (4–5 marks)

4 of 6 models expect this — consensus probability 0.69

Non-right-angled trigonometry is the panel's top topic (97% examined), and 2025 was the first paper since 2019 without a bearings or radial-survey question — four models call it back: two legs on true bearings, cosine rule for the joining distance, sine rule for an angle, then the killer final part, the bearing of one endpoint from the other. The "bearing of A from B" language is flagged in the feedback of every year the format ran (2020, 2022, 2023, 2024).

Practise

always redraw the diagram and transfer north lines through every vertex before touching the calculator.

Try it → Q39 4 of 6 models expect this — see the evidence

4. The interest-factor table (annuities)

4 of 6 models expect this — consensus probability 0.66

Four models, 0.66 — and the multiply-vs-divide error is flagged in every feedback document 2019–2025, the most reliable trap in the subject. The rate is quoted per annum but contributions are monthly or quarterly: convert $r$ and $n$ before choosing the factor, then divide the target by the factor to get the contribution (multiply only when you want a future value).

Practise

say out loud what the factor means ("the future value of $1 per period") before deciding which way to use it.

Try it → Q27 4 of 6 models expect this — see the evidence

5. Regression with the interpretation marks (3–6 marks)

6 of 6 models expect this — consensus probability 0.64

The only cluster all six models agree on (64% chance), on the 2025 Q25 template: gradient and intercept read off the graph or given, both interpreted in context (units and direction, not a textbook definition), one interpolation, then one mark for explaining why an extrapolated prediction fails — a negative or impossible value, or simply outside the data. Generic instead of contextual interpretation is the single most-repeated feedback theme (2021, 2023, 2025).

Practise

write slope interpretations as "for each extra [unit of x], [y] changes by [gradient] [units]".

Try it → Q25 6 of 6 models expect this — see the evidence

Strong candidates

  • Appliance running cost (67% chance) — watts→kilowatts, cents→dollars, and increasingly the reverse: given a seasonal total, recover daily hours of use (the 2025 air-conditioner lineage).
  • Depreciation both ways (66% chance) — declining-balance vs straight-line on the same asset, or deduce the hidden rate from a table of salvage values; know that "depreciation during year $n$" is a subtraction of two values.
  • Composite solids (65% chance) — hemisphere + cylinder/cone: halve the reference-sheet formulae, convert diameters to radii, drop the joined faces, then convert cm³ to litres. Flagged in 2020, 2022 and 2024.
  • The z-table chain (62% chance) — the printed $P(Z \lt z)$ table is due back after its 2025 rest: non-integer z-score → table → complement → expected count (a whole number, not a probability).
  • Two draws without replacement (62% chance) — tree diagram, denominators drop by one, multiply along branches, add the favourable paths.
  • BAC or a supplied formula (63% chance) — substitute, rearrange, then convert decimal hours to hours-and-minutes for a clock-time answer. 0.75 hours is 45 minutes, not 75.
  • Trapezoidal rule, twice, from a scale plan (57% chance) — convert the scale before applying the rule, then $V = Ah$; a mark rides on rounding to the stated significant figures.
  • Income tax table (61% chance) — but see the dissent below.

A dissent worth knowing, and streaks that can break

One model rates the income-tax table only 0.35, on a sharp pattern read: the tax table has run in odd years only (2021, 2023, 2025) and skipped 2022 and 2024 — so 2026 may examine money matters through earnings, commission and GST instead. Prepare the tax table anyway (four models still back it at 0.6–0.8), but don't be shocked by its absence. The 2025 backtest of this method in chemistry and maths ext 1 is the broader caution: two Ext 1 predictions carried by five of six models simply didn't appear. Streaks break — expect this paper to break one somewhere too.

The honest fine print

Backtesting this method against the real 2025 papers (chemistry and maths ext 1): every topic rated ≥90% appeared (37/37 across both pilot subjects), roughly half of the specific question predictions recognisably appeared, and the misses clustered where the examiners twisted a format or moved a question into multiple choice. The lesson for revision: master the skill chains and setting-out habits above — unit conversions done one step at a time, diagrams redrawn, justification sentences written — because they survive every format twist; a memorised question doesn't.

Want to check our working? every call above, with each model's own prediction
5 of 6 models expect this — consensus probability 0.70 Quadratic model in context: maximum via the axis of symmetry 5 of 6 models expect this
std2-q1-quadratic-model Section II 3–5 marks mid range consensus 0.70

A supplied parabola modelling revenue, enclosure area, braking distance or a projectile: locate the turning point from the axis of symmetry (or given intercepts), evaluate the maximum, then either find the second input giving an equal output or justify the sensible practical domain (only the positive branch is valid). Vertex-vs-intercept confusion and ignoring the practical domain are flagged in the 2021, 2024 and 2025 feedback. The panel's strongest question-level consensus.

What each model said

  • Claude Fable 5 0.80
  • Gemini 3.1 Pro 0.75
  • Claude Opus 5 0.72
  • DeepSeek V4 0.70
  • Grok 4.6 0.55

In the practice paper: Q22

5 of 6 models expect this — consensus probability 0.69 Minimum spanning tree from a distance table, plus a shortest-path perturbation 5 of 6 models expect this
std2-q2-mst-network Section II 3–5 marks mid range consensus 0.69

A distance/cable table between five to eight towns is first drawn as a weighted network, the MST is drawn on a supplied blank diagram and its total length stated; the follow-up perturbs the network — a road closure forces the new shortest path (every vertex listed), or the question asks whether an equal-weight second tree exists. The designed trap in every lineage year: the shortest path does NOT lie inside the MST (2019 Q30, 2023 Q19, 2024 Q16). Fable carries the same call as a topic-level bold call (MST returns after its first absence in seven years, 0.5) rather than a question row.

What each model said

  • Gemini 3.1 Pro 0.85
  • DeepSeek V4 0.70
  • Grok 4.6 0.70
  • Claude Opus 5 0.60
  • GPT-5.6 Sol 0.58

In the practice paper: Q19

4 of 6 models expect this — consensus probability 0.69 Two-leg bearings navigation: cosine rule, sine rule, then 'bearing of A from B' 4 of 6 models expect this
std2-q3-bearings-navigation Section II 4–5 marks discriminator consensus 0.69

Two legs from a common point on given true bearings: cosine rule for the joining distance, sine rule for an angle, and the final part asks the bearing of one endpoint from the other — returning after its 2025 absence (2020 Q31 / 2022 Q33 template). DeepSeek frames the same chain inside a radial survey. 'Bearing of A from B' language and the parallel-line angle transfer are flagged in the feedback of every year the format ran (2020, 2022, 2023, 2024). MS-M6 is the panel's top-rated topic (P(examined) 0.97).

What each model said

  • Gemini 3.1 Pro 0.85
  • DeepSeek V4 0.70
  • Claude Fable 5 0.65
  • GPT-5.6 Sol 0.56

In the practice paper: Q39

4 of 6 models expect this — consensus probability 0.69 Critical path analysis: EST/LST scan, float, and a duration-change justification 4 of 6 models expect this
std2-q4-critical-path Section II 4–5 marks mid range consensus 0.69

Activity table or network with durations: forward/backward scan, state the critical path and minimum completion time, then compute $\text{float} = \text{LST} - \text{EST}$ for a named non-critical activity and justify whether a changed duration alters the critical path (2025 Q19(c) style). Opus adds the crashing variant (which single activity to shorten). NOTE: MS-N3 is the panel's one contested topic (stddev 0.259, two rested votes — gemini and sol carry no N3 question rows); the four models that did predict it agree closely.

What each model said

  • Claude Opus 5 0.72
  • DeepSeek V4 0.70
  • Claude Fable 5 0.70
  • Grok 4.6 0.62

In the practice paper: Q37

4 of 6 models expect this — consensus probability 0.67 Appliance running cost: W→kW and cents→dollars, possibly run in reverse 4 of 6 models expect this
std2-q5-running-cost Section II 2–4 marks mid range consensus 0.67

Given power in watts and a tariff in cents per kWh, compute the cost of running an appliance over stated hours — the 650 W air conditioner of 2025 Q21 is the direct lineage — or the reverse: given a seasonal total cost, recover the average daily hours of use. DeepSeek's variant compares a heater with a heat pump. The kW/W and cents/dollars conversions each effectively carry a mark (flagged 2021, 2024, 2025).

What each model said

  • Claude Fable 5 0.75
  • Claude Opus 5 0.68
  • GPT-5.6 Sol 0.64
  • DeepSeek V4 0.60

In the practice paper: Q17

4 of 6 models expect this — consensus probability 0.66 Annuity interest-factor table with the rate/period conversion trap 4 of 6 models expect this
std2-q6-annuity-factor-table Section II 2–5 marks mid range consensus 0.66

A future-value (or present-value) of \$1 table where the stated rate is per annum but contributions are monthly/quarterly: convert r and n BEFORE selecting the factor, then divide the target by the factor for the contribution or repayment (multiply for a future value). The multiply-vs-divide error is flagged in every feedback document 2019–2025 — the most reliable trap in the subject. Gemini holds the present-value form (repayment on a \$400,000 loan).

What each model said

  • Claude Fable 5 0.75
  • Gemini 3.1 Pro 0.70
  • Claude Opus 5 0.70
  • Grok 4.6 0.50

In the practice paper: Q27

4 of 6 models expect this — consensus probability 0.66 Declining-balance vs straight-line depreciation — compare, or recover the hidden rate 4 of 6 models expect this
std2-q7-depreciation-comparison Section II 3–5 marks mid range consensus 0.66

One asset, both methods: compute both book values over the same term and state which method leaves the higher/lower salvage value — or (fable) the rate is hidden and must be deduced from successive salvage values in a table before projecting, with the 'depreciation DURING the nth year' subtraction as the discriminator step. Salvage-value vs total-depreciation confusion flagged 2022, 2023, 2024, 2025.

What each model said

  • Gemini 3.1 Pro 0.80
  • Claude Opus 5 0.68
  • Claude Fable 5 0.60
  • Grok 4.6 0.55

In the practice paper: Q31

4 of 6 models expect this — consensus probability 0.65 Composite solid (hemisphere + cylinder/cone/pyramid): surface area or volume 4 of 6 models expect this
std2-q8-composite-solid Section II 3–5 marks discriminator consensus 0.65

A hemisphere joined to a cylinder, cone, prism or spheres packed in a pyramid (2025 Q26(b) / 2024 Q34 lineage): halve or combine reference-sheet formulae, convert diameters to radii, exclude the joined faces from surface area, then convert units (cm³→L or cm²→m²) — with fable and opus adding a reverse step (solve for a height from a given empty-space volume). Diameter/radius and area-vs-volume formula confusion flagged 2020, 2022, 2024.

What each model said

  • Claude Fable 5 0.75
  • Claude Opus 5 0.75
  • GPT-5.6 Sol 0.59
  • Grok 4.6 0.50

In the practice paper: Q26

6 of 6 models expect this — consensus probability 0.64 Least-squares regression: slope and intercept IN CONTEXT, interpolate, then critique extrapolation 6 of 6 models expect this
std2-q9-regression-interpretation Section II 3–6 marks mid range consensus 0.64

The panel's only unanimous cluster. A scatterplot with the LSRL drawn (read gradient from two grid points and the intercept) or the equation supplied: interpret BOTH slope and y-intercept in context, predict for an x inside the data range, then a one-mark part explaining why an extrapolated prediction fails (negative/impossible value or outside the data) — the 2025 Q25 four-part template. Generic rather than contextual interpretation is the single most-repeated feedback theme (2021, 2023, 2025).

What each model said

  • Gemini 3.1 Pro 0.80
  • DeepSeek V4 0.70
  • Claude Fable 5 0.70
  • Claude Opus 5 0.66
  • Grok 4.6 0.52
  • GPT-5.6 Sol 0.46

In the practice paper: Q25

4 of 6 models expect this — consensus probability 0.63 BAC / supplied-formula substitution chained to a time and a clock answer 4 of 6 models expect this
std2-q10-bac-formula Section II 2–4 marks discriminator consensus 0.63

Blood alcohol content returns after its 2025 absence: the formula is printed; convert glasses to standard drinks, substitute $N$, $H$, $M$, then chain to $\text{Time} = \text{BAC}/0.015$ and a clock time to the nearest minute — or the 2023 Q36 reverse, solving for $H$. Gemini and sol generalise to any supplied formula (medication dosage, workplace). The decimal-hours-to-minutes conversion is the flagged killer (2019, 2023, 2024). DeepSeek separately carries Young's-formula dosage as a bold call (unexamined since 2019, 0.35).

What each model said

  • Gemini 3.1 Pro 0.75
  • DeepSeek V4 0.70
  • Claude Fable 5 0.55
  • GPT-5.6 Sol 0.53

In the practice paper: Q30

5 of 6 models expect this — consensus probability 0.62 Normal distribution: printed z-table → tail probability → expected count 5 of 6 models expect this
std2-q11-ztable-expected-count Section II 3–4 marks discriminator consensus 0.62

The printed $P(Z \lt z)$ table returns after its 2025 absence: compute a non-integer z-score from $\mu$ and $\sigma$ (birth weights, koala masses, exam scores), read the table, complement or use symmetry for the required tail, then multiply by a population count and report a whole-number expected count (2021 Q38 / 2023 Q38 / 2024 Q35 template). Grok's reconstruction $P(Z>0.3)=0.3821$ is verbatim from the pack. 'A z-score is not a probability' is flagged every year 2021–2025.

What each model said

  • Claude Fable 5 0.65
  • Grok 4.6 0.65
  • GPT-5.6 Sol 0.62
  • DeepSeek V4 0.60
  • Claude Opus 5 0.60

In the practice paper: Q36

5 of 6 models expect this — consensus probability 0.62 Two-stage probability without replacement: tree diagram, $P(\text{same colour})$ 5 of 6 models expect this
std2-q12-without-replacement Section II 2–4 marks mid range consensus 0.62

Two selections from a small mixed set (marbles, chocolates, raffle tickets — the 2019 apples / 2023 raffle / 2025 lollies template with fresh objects): tree diagram with reduced second-stage denominators, multiply along branches, add the mutually exclusive favourable paths for $P(\text{same colour})$ or $P(\text{at least one})$. Add-vs-multiply and unadjusted denominators flagged 2019, 2023, 2025.

What each model said

  • Gemini 3.1 Pro 0.70
  • Claude Fable 5 0.65
  • Claude Opus 5 0.62
  • GPT-5.6 Sol 0.60
  • DeepSeek V4 0.55

In the practice paper: Q18

5 of 6 models expect this — consensus probability 0.61 Progressive income-tax table + Medicare levy (forward, or solved in reverse) 5 of 6 models expect this
std2-q13-income-tax Section II 3–4 marks mid range consensus 0.61

Gross salary minus deductions gives taxable income; apply the correct bracket (cents vs dollars trap), add the 2% Medicare levy — or the opus/fable reverse: given the tax payable, recover the taxable income within one bracket. Bracket selection and taxable-income vs tax-payable confusion flagged 2019–2025. Dissent worth knowing: fable rates this only 0.35, holding a contrarian bold call (0.4) that the strict odd-years-only tax-table pattern (2021/2023/2025) means NO tax table in 2026.

What each model said

  • Gemini 3.1 Pro 0.80
  • Claude Opus 5 0.68
  • DeepSeek V4 0.60
  • Grok 4.6 0.60
  • Claude Fable 5 0.35

In the practice paper: Q21

5 of 6 models expect this — consensus probability 0.57 Trapezoidal rule applied twice, from a scaled plan, feeding a volume 5 of 6 models expect this
std2-q14-trapezoidal-scale Section II 3–6 marks discriminator consensus 0.57

Two applications of the trapezoidal rule over offsets read from a scaled site plan, map or dimensioned cross-section: convert the scale BEFORE applying the rule, then feed the area into $V = Ah$ for a rainfall volume, soil to remove, or a material cost, with a mark attached to rounding to stated significant figures. Sol holds the scale-plan→area→cost variant without naming the rule. 'Convert scale lengths before the trapezoidal rule' flagged 2019, 2020, 2023, 2025.

What each model said

  • Claude Fable 5 0.70
  • DeepSeek V4 0.60
  • GPT-5.6 Sol 0.58
  • Claude Opus 5 0.52
  • Grok 4.6 0.45

In the practice paper: Q33

Watch list — worth having ready (13)

  • Two triangles sharing a side — right-angled trig for the link, sine/cosine rule for the second unknown, incl. the flagpole-on-sloping-ground setting, which IS in the pack (4 models, ~0.61)

  • Time zones with a flight duration: answer must state BOTH clock time and day; grok's Sydney→New York 8:20 pm / 20 h 24 min is a verbatim 2024 Q37 reconstruction (5 models, ~0.55)

  • Credit-card interest, daily compounding around an interest-free period — day-count wording is the trap (4 models, ~0.57)

  • z-score relative-performance comparison across subjects, usually MC — raw-mark comparison is the distractor (4 models, ~0.57)

  • Box-plot from an ogive or frequency table + outlier justification via $Q_1 - 1.5 \times \text{IQR}$ with a stated conclusion (4 models, ~0.55)

  • Inverse variation: find $k$ from one pair, complete a table, sketch the hyperbola — the 2023 painters question is the pack lineage (4 models, ~0.55)

  • Empirical-rule non-standard fractions (84%/16%/97.5%/2.5%) run in reverse to a raw score, mean or sd (3 models, ~0.55)

  • Radial survey with a triangle area given: rearrange $A = \tfrac{1}{2}ab\sin C$ for the included angle, then a bearing (3 models, ~0.55)

  • Two-phase annuity as the paper's closer: PV table, withdrawals change amount partway, defer the second phase (5 models, ~0.50 — the panel's widest-held watch item)

  • Break-even graphical simultaneous equations: two linear plans, intersection, 'which is cheaper AND by how much' (4 models, ~0.52)

  • Max-flow min-cut: cut capacity counting only source→sink edges, which pipe to increase (4 models, ~0.49)

  • Change-the-subject MC — fable's 0.8 is the panel's single highest question probability (2019/2022/2024/2025 MC lineage, 1 model)

  • GST reverse reasoning on a receipt (×10 vs ÷11 trap); grok's $124.87/$3.86 is a verbatim pack reconstruction (3 models, ~0.45)

Where to spend your study time

How likely each topic is to appear this year.

Non-right-angled trigonometry 97% likely

Chance of a big question (4+ marks) here: 84%

Question types predicted here extended response ×8 short answer ×5 multiple choice ×2

What each model expects

  • DeepSeek V4: A 5-mark container radial survey problem (like 2021 Q39) including ambiguous case for an angle, testing obtuse angle recognition.
  • Claude Fable 5: A 4-5 mark bearings question with cosine-then-sine-rule chaining and a final 'bearing of ... from ...' part sits in the last third of Section II.
  • Gemini 3.1 Pro: A heavy 5-mark question chaining a compass bearing interpretation into the cosine rule.
  • GPT-5.6 Sol: One substantial navigation problem will chain a bearing angle, a non-right-triangle rule and a final bearing interpretation.
  • Grok 4.6: Mid Section II will carry a 3–4 mark radial-survey or two-triangle bearings question, plus one Section I true-bearing conversion.
  • Claude Opus 5: A compass radial survey returns after a two-year absence, with the area-of-a-triangle rule rearranged for an included angle and the answer demanded as a three-figure bearing to the nearest degree.
Types of relationships (linear, quadratic, reciprocal, exponential models) 96% likely

Chance of a big question (4+ marks) here: 79%

Question types predicted here extended response ×10 short answer ×4 multiple choice ×1

What each model expects

  • DeepSeek V4: A quadratic optimisation question in Section II Q25–Q30, worth 4 marks, asking to find max revenue from a demand model and interpret the axis of symmetry.
  • Claude Fable 5: A quadratic-maximum question using the axis of symmetry appears again, and the break-even graphical simultaneous solve returns after being reduced to equation-writing in 2025.
  • Gemini 3.1 Pro: A 4-mark quadratic modelling question will ask students to justify why the model fails outside a specific domain.
  • GPT-5.6 Sol: The substantial A4 question will shift from another routine parabola towards a reciprocal or competing-linear model.
  • Grok 4.6: Section II will include a graphed quadratic with a practical-domain part, and either a two-provider break-even or an inverse-variation table-and-sketch.
  • Claude Opus 5: A supplied parabola is mined twice — once for the turning point and once for a second value obtained purely by symmetry — with no algebraic solving demanded, placed in the first half of Section II.
Investments and loans 96% likely

Chance of a big question (4+ marks) here: 77%

Question types predicted here short answer ×9 extended response ×5 multiple choice ×1

What each model expects

  • DeepSeek V4: A 4-mark credit card or reducing balance loan question with spreadsheet completion, requiring interpretation of interest charged and explanation of why balance never decreases.
  • Claude Fable 5: A depreciation question that hides the rate behind a table or graph, targeting depreciation as a loss amount rather than salvage value, sits in the second half of Section II.
  • Gemini 3.1 Pro: A multi-part credit card interest calculation will appear in the mid-30s of Section II.
  • GPT-5.6 Sol: F4 will still appear, but its main task will be a shorter reducing-balance or compound-rate problem after heavy 2025 coverage.
  • Grok 4.6: A 3–4 mark compare-two-depreciation or reducing-balance spreadsheet will sit in mid Section II; credit-card daily compounding is less likely after 2025 Q27.
  • Claude Opus 5: Depreciation appears once and is deliberately hybrid — two methods on one asset, or a rate recovered from a supplied salvage-value graph — with the answer being an amount of depreciation rather than a salvage value.
The normal distribution 94% likely

Chance of a big question (4+ marks) here: 72%

Question types predicted here extended response ×6 multiple choice ×4 short answer ×3

What each model expects

  • DeepSeek V4: A 4-mark question with two parts: (a) find z-score and probability; (b) use given table to find cut-off score (inverse normal) for a stated percentage.
  • Claude Fable 5: A supplied z-table question worth 3-4 marks returns, requiring a complement step and an expected-count multiplication, after 2025 ran empirical-rule-only.
  • Gemini 3.1 Pro: A comparison of two datasets using z-scores to justify which result is relatively better.
  • GPT-5.6 Sol: A three- or four-mark tail-probability question will require converting a table value into an expected number of observations.
  • Grok 4.6: A 4-mark table-based z-score count (not empirical-only) will appear late in Section II, reversing 2025’s empirical-heavy Q40.
  • Claude Opus 5: Normal distribution takes exactly one multiple-choice mark plus a single 4–5 mark Section II question sited in the last five questions of the paper, with a z-probability table supplied.
Data analysis, relative frequency and probability (Y11) 93% likely

Chance of a big question (4+ marks) here: 52%

Question types predicted here short answer ×11 extended response ×2 multiple choice ×1

What each model expects

  • DeepSeek V4: A box-plot construction and interpretation question worth 4 marks, including an outlier identification step, placed early in Section II.
  • Claude Fable 5: The without-replacement 'same colour' tree-diagram question reappears with a new context, worth 2-3 marks in mid-Section II.
  • Gemini 3.1 Pro: A multi-stage probability question without replacement using a tree diagram.
  • GPT-5.6 Sol: Probability without replacement will return, accompanied by at least one data-representation or sampling item.
  • Grok 4.6: Either an outlier-justification or a parallel box-plot comparison, and a realistic chance of Pareto returning after a three-year gap.
  • Claude Opus 5: Probability is confined to a single 2–3 mark two-stage without-replacement item plus one or two multiple-choice questions on relative frequency, sampling method or expected frequency.
Annuities 93% likely

Chance of a big question (4+ marks) here: 69%

Question types predicted here extended response ×9 short answer ×4

What each model expects

  • DeepSeek V4: A 4–5 mark question requiring the use of a present value or future value table with multiple cash flow streams, mixing standard annuity with a lump sum or changing amounts.
  • Claude Fable 5: The final or penultimate question is an interest-factor-table annuity synthesis with a changed payment phase, echoing 2020 Q37 and 2024 Q41.
  • Gemini 3.1 Pro: A 5-mark question involving a table of interest factors to find a loan repayment amount.
  • GPT-5.6 Sol: The annuity task will use either a present-value withdrawal stream or recurrence relation rather than another direct future-value multiplication.
  • Grok 4.6: Annuities will again be table-driven; the discriminator is a two-stage PV or a rate-change after n years, likely among the last three questions.
  • Claude Opus 5: At least one annuity question supplies an interest-factor table whose column cannot be used until the annual rate is divided by the compounding frequency, and the mark scheme rewards identifying the correct factor separately from the arithmetic.
Bivariate data analysis 92% likely

Chance of a big question (4+ marks) here: 71%

Question types predicted here short answer ×6 extended response ×5 multiple choice ×3

What each model expects

  • DeepSeek V4: A 4-mark question in Section II (likely Q28–Q30) combining plotting, line of best fit, interpretation, and a 'what if' scenario with a new data point.
  • Claude Fable 5: Gradient-and-intercept interpretation in context recurs, with a reserved one-mark extrapolation-limitation explanation, in a 4-6 mark multi-part bivariate question.
  • Gemini 3.1 Pro: Students will calculate Pearson's correlation coefficient and explicitly distinguish it from data skewness.
  • GPT-5.6 Sol: Following the six-mark 2025 task, bivariate analysis will contract to a short interpretation item or multiple choice rather than another major discriminator.
  • Grok 4.6: A 4–5 mark scatterplot with contextual slope interpretation and an extrapolation caveat, lighter than 2025’s 6-mark Q25 but still present.
  • Claude Opus 5: Bivariate data carries the paper's longest single question (5–6 marks) as a chained item: describe form and direction, interpret slope and intercept in context, predict inside the range, then reject an extrapolation.
Network concepts (paths, trees, shortest path) 92% likely may rest

Chance of a big question (4+ marks) here: 67%

Question types predicted here short answer ×5 extended response ×4 multiple choice ×1

What each model expects

  • DeepSeek V4: A 3-mark MST question from a distance table, directly after a multiple choice on network degree or edges, ensuring Network Concepts are assessed both ways.
  • Gemini 3.1 Pro: A 3-mark minimum spanning tree question applied to a utilities/infrastructure scenario.
  • GPT-5.6 Sol: After only two multiple-choice marks in 2025, N2 will return as a four- or five-mark Section II minimum-spanning-tree problem.
  • Grok 4.6: Section II will require drawing an MST from a table or pre-drawn network and stating its length; 2025 only tested MST as MC, so a 4–5 mark tree is due.
  • Claude Opus 5: MS-N2 returns to Section II after being confined to multiple choice in 2025, with a minimum spanning tree built from a supplied distance table or map and worth 3–4 marks.
Money matters (Y11) 89% likely

Chance of a big question (4+ marks) here: 51%

Question types predicted here short answer ×12 extended response ×1

What each model expects

  • DeepSeek V4: A 3-mark tax or simple interest question, possibly involving GST component extraction, continuing the trend of practical financial literacy.
  • Claude Fable 5: Money Matters is examined through earnings and GST rather than a full income-tax table, which has run only in odd years (2021, 2023, 2025) since 2020.
  • Gemini 3.1 Pro: A standard 4-mark income tax calculation involving allowable deductions and the Medicare levy.
  • GPT-5.6 Sol: F1 will move away from another five-mark car loan and use wages, commission or GST as its main Section II application.
  • Grok 4.6: A forward tax-plus-Medicare 3-mark question, reversing 2025’s find-income-from-tax Q31, plus possibly a GST-free remainder.
  • Claude Opus 5: An income-tax table question appears once, worth 3 marks, and is set as an inverse problem — tax payable or a Medicare levy is given and the taxable income must be recovered.
Formulae, equations, linear relationships (Y11) 89% likely may rest

Chance of a big question (4+ marks) here: 42%

Question types predicted here short answer ×6 multiple choice ×1 extended response ×1

What each model expects

  • DeepSeek V4: A 3-mark BAC question in Section II, using the given formula and then the time formula, with conversion to hours and minutes.
  • Claude Fable 5: A blood-alcohol-content question reappears in Section II with the formulae printed, most likely the chained time-to-zero-then-clock-time form.
  • Gemini 3.1 Pro: A question requiring algebraic manipulation of a given practical formula, testing substitution and rounding skills.
  • GPT-5.6 Sol: A supplied-formula problem will require solving for a non-subject variable and completing a practical time or dosage interpretation.
Rates and ratios (incl. scale drawings, energy) 87% likely may rest

Chance of a big question (4+ marks) here: 72%

Question types predicted here extended response ×6 short answer ×5 multiple choice ×2

What each model expects

  • DeepSeek V4: A 3–4 mark energy efficiency question, possibly linked to a new appliance comparison, testing conversion of kW and hours to kWh and cost.
  • Claude Fable 5: An electricity-cost question requiring a W-to-kW conversion appears for the fourth time in six years, worth 2-3 marks in the first half of Section II.
  • GPT-5.6 Sol: Section II will contain a four- or five-mark scale-and-rate chain ending in a practical cost, capacity or material decision.
  • Grok 4.6: A scale-plus-trapezoidal (possibly with rainfall volume) remains the likely 5–6 mark measurement discriminator; capture-recapture should be rested after 2025 Q24.
  • Claude Opus 5: An energy-cost question runs in both directions within one item — cost from power and time in part (a), then hours from a total cost in part (b) — as in 2025 Q21 but with the tariff or appliance changed.
Applications of measurement, working with time (Y11) 85% likely may rest

Chance of a big question (4+ marks) here: 66%

Question types predicted here short answer ×10 extended response ×3

What each model expects

  • DeepSeek V4: A trapezoidal rule question with subsequent volume calculation (like 2025 Q26), appearing in the first half of Section II.
  • Claude Fable 5: A two-application trapezoidal-rule question chained to a volume-and-unit-conversion part, worth 4-6 marks across a multi-part question in mid-Section II.
  • GPT-5.6 Sol: A late-middle Section II measurement problem will combine composite geometry, excluded surfaces and a unit conversion.
  • Grok 4.6: A 3-mark time-zone flight (day and time) is due after 2025’s MC-only Q12, plus a 4-mark composite surface-area.
  • Claude Opus 5: A composite-solid surface area or capacity question appears in the second half of Section II and is worth 3–4 marks, with a unit conversion (litres, or square metres) built into the final line.
Critical path analysis, flow networks 77% likely models split may rest

Chance of a big question (4+ marks) here: 69%

Question types predicted here extended response ×6 short answer ×3

What each model expects

  • DeepSeek V4: A 5-mark CPA question, possibly combining with a flow network sub-question to cover both N3 subtopics in one question.
  • Claude Fable 5: Both a critical-path question with a duration-change justification and a flow question appear again, trimmed to roughly 7-8 combined marks after the 10-mark 2025 allocation.
  • Grok 4.6: A standard EST/LST critical-path question with one float calculation is more likely than another 5-mark flow network after 2025 Q22.
  • Claude Opus 5: Critical path analysis is set once, worth 4–5 marks, and includes a one-mark 'does this change the critical path? give a reason' part answered by comparing an increase against a float time.

How likely each topic is to appear. Open a topic for the question types to practise there.

Practise with the paper

Maths Standard 2 practice paper

100 marks · 40 questions

Every question is traceable to the consensus prediction behind it — open the web version and each question carries a “why this question” link into the evidence. All questions are original Intuition compositions in NESA style.

Read it on the web Download PDF Solutions (PDF)

Exam Success Program

Book our ESPs now — online or in person

A paper on your desk is one thing — the clock running is another. Sit Maths Standard 2 practice exams under real conditions and get them marked with real feedback. Exam times run seven days a week, right up until the HSC begins.

Book your ESP →

Places are still open for 2026

Intu AI

One paper isn't enough? Generate more

Intu AI builds unlimited practice questions for Maths Standard 2 in these styles, marks your working, and explains what you missed — aligned to your syllabus.

Practise with Intu AI →

Share & save

Practice paper (PDF)

Every style card and paper question has its own link — hover any card and use its copy-link icon to share exactly the thing you mean. Printing this page gives a clean copy too.

Published Aug 2026, before the exams. In November 2026 we score these predictions publicly against the real paper — per-model calibration and question-level hit rates, the same harness as the 2025 backtest. How we did it.