In our latest benchmark, Silent Student answered 156 of 164 graded answers correctly (95%) across seven homework assignments built as replicas of WebAssign, Cengage MindTap, McGraw Hill Connect, Pearson MyLab and Macmillan Achieve, and submitted six of the seven. It ran on the same engine the Base plan uses. The one it didn’t submit is in the Macmillan Achieve AI demo.
Results by assignment
| Assignment | Replica modeled on | What it tests | Correct | Submitted |
|---|---|---|---|---|
| STEM homework workbook | WebAssign | Per-part answer checks, retries, figures and a required submission note | 8 of 8 | Yes |
| Mixed assessment | Cengage MindTap | Two-pane activity player, custom pickers, one attempt per question | 9 of 9 | Yes |
| Accounting workbook | McGraw Hill Connect | Dense multi-cell grids, requirement tabs and a question map, 109 graded parts | 102 of 109 | Yes |
| Calculus homework | Cengage WebAssign | Graphs as answer choices, 3D figures, a 10-part accordion, math notation | 16 of 16 | Yes |
| Finance homework (first third) | Pearson MyLab | Custom pickers, per-part checks, three tries per part | 9 of 9 | Yes |
| Statistics assignment (first third) | Macmillan Achieve | Sorting bins, numeric and short answers, image choice | 7 of 8 | No — stopped on one question |
| College algebra (first third) | Pearson MyLab (College Algebra) | Math editor inside nested frames, immediate feedback | 5 of 5 | Yes |
| Total | 156 of 164 (95.1%) | 6 of 7 |
Watch the runs
These seven clips are separate recorded runs of the same replicas, made on September 28 and 29, 2026, after the scored run on September 22, 2026 that the table above reports. Each is sped up so the progress is easy to follow and shows its own read-back result, which can differ from the table because every run of the agent makes its own choices; where it differs, the caption says so. Each clip is a replica we built for testing, modeled on the named platform, not the real site, and no platform maker is affiliated with or has endorsed them. The agent is the Base plan’s vision engine, run in a standalone test harness, not the full desktop app. The on-screen label “Nate’s Empire” is the benchmark’s name.
STEM homework workbook · recorded September 28, 2026
Replica modeled on WebAssign
8 of 8 graded parts correct Submitted
The agent works a WebAssign-style STEM workbook: numeric answers, multi-blank answers and short written answers, several of them read off figures. It checks each part as it goes, writes the required submission note and submits.
A benchmark replica we built for testing, modeled on WebAssign. It is not the real site, and the platform's maker is not affiliated with it and has not endorsed it. The agent is the Base plan's vision engine, run in a standalone test harness, not the full desktop app.
Mixed assessment · recorded September 28, 2026
Replica modeled on Cengage MindTap
9 of 9 graded parts correct Submitted
The agent works a MindTap-style two-pane activity player: multiple choice, true/false, matching, select-all, dropdown, short-answer and fill-in-the-blank parts, one attempt per question, then submits.
A benchmark replica we built for testing, modeled on Cengage MindTap. It is not the real site, and the platform's maker is not affiliated with it and has not endorsed it. The agent is the Base plan's vision engine, run in a standalone test harness, not the full desktop app.
Accounting workbook · recorded September 29, 2026
Replica modeled on McGraw Hill Connect
106 of 109 graded parts correct Submitted
This recording is a separate run with a different result from the benchmark table: the scored run on September 22, 2026 got 102 of 109, submitted; this recording got 106 of 109, submitted.
The agent works a Connect-style accounting workbook of integrated business models: dense grids of numeric cells and dropdowns spread across requirement tabs, with a question map for moving between items. It checks its work as it goes, fixes cells until they check out, and submits.
A benchmark replica we built for testing, modeled on McGraw Hill Connect. It is not the real site, and the platform's maker is not affiliated with it and has not endorsed it. The agent is the Base plan's vision engine, run in a standalone test harness, not the full desktop app.
Calculus homework · recorded September 29, 2026
Replica modeled on Cengage WebAssign
16 of 16 graded parts correct Submitted
The agent works a WebAssign-style calculus homework on volumes of solids: it picks graphs and 3D figures as answer choices, opens each part of a long accordion question, types answers in math notation, and submits.
A benchmark replica we built for testing, modeled on Cengage WebAssign. It is not the real site, and the platform's maker is not affiliated with it and has not endorsed it. The agent is the Base plan's vision engine, run in a standalone test harness, not the full desktop app.
Finance homework (first third) · recorded September 29, 2026
Replica modeled on Pearson MyLab Finance
9 of 9 graded parts correct Submitted
The agent works the first third of a MyLab-style finance homework: numeric answers, dropdown pickers and a multiple-choice part, checking each part before moving on, then submits.
A benchmark replica we built for testing, modeled on Pearson MyLab Finance. It is not the real site, and the platform's maker is not affiliated with it and has not endorsed it. The agent is the Base plan's vision engine, run in a standalone test harness, not the full desktop app.
Statistics assignment (first third) · recorded September 29, 2026
Replica modeled on Macmillan Achieve
6 of 8 graded parts correct Not submitted
This recording is a separate run with a different result from the benchmark table: the scored run on September 22, 2026 got 7 of 8, not submitted; this recording got 6 of 8, not submitted.
The agent works the first third of an Achieve-style statistics assignment: box plots, numeric answers, dropdowns and a five-number summary. Partway through it stops making progress; the run ended on its idle limit and the assignment was never submitted. The clip keeps that stretch instead of cutting it.
A benchmark replica we built for testing, modeled on Macmillan Achieve. It is not the real site, and the platform's maker is not affiliated with it and has not endorsed it. The agent is the Base plan's vision engine, run in a standalone test harness, not the full desktop app.
College algebra (first third) · recorded September 29, 2026
Replica modeled on Pearson MyLab College Algebra
5 of 5 graded parts correct Submitted
The agent works the first third of a MyLab-style college algebra homework, typing each answer into a math editor inside nested frames, reading the immediate feedback, and submits.
A benchmark replica we built for testing, modeled on Pearson MyLab College Algebra. It is not the real site, and the platform's maker is not affiliated with it and has not endorsed it. The agent is the Base plan's vision engine, run in a standalone test harness, not the full desktop app.
How the benchmark works
- Replicas, not live accounts. Each assignment is a local replica built to behave like the real platform: the same question types, answer controls, checks and submission flow. No real student account or live vendor site was used.
- Deliberate faults. Four of the assignments sometimes reject a correct answer or make a control disappear, to test whether the bot recovers instead of giving up. It recovered from 9 of the 12.
- Scored from the replica. After submission, each replica is read back to see which answers it recorded as correct. The bot’s own report is not used.
- Same engine as the app. The run used the Vision agent with the model and settings the Base plan uses.
- When. One complete run of all seven assignments on September 22, 2026.
What this doesn’t show
It is one run of one engine on seven assignments. Your course, question types and platform version may differ, live platforms change, and some question types are harder than others. It says nothing about proctored exams, which Silent Student does not take.
Has anyone tested Silent Student independently?
Not yet. As of October 19, 2026 the only published accuracy test is this one, run by the company on replicas it built. The method, per-assignment results and recordings are here so anyone can check them. See is Silent Student legit? for what else you can check.
Were these real WebAssign, Pearson or McGraw Hill accounts?
No. Each assignment is a replica we built for testing, so no real student accounts or live vendor sites were used. The Pearson Finance and Achieve replicas are authorized one-to-one captures of the first questions of real assignments; the others are clean-room reconstructions of how those platforms behave.
How was each answer scored?
After each assignment was submitted, the replica itself was read back to see which answers it recorded as correct. The score does not rely on the bot's own report.
Does this benchmark predict my results?
It measures one engine on seven assignments. Your course, question types and platform version may differ, and some question types are harder than others.
Last updated October 19, 2026. See also security and privacy, supported platforms, pricing, comparisons with other tools, the FAQ and the kinds of homework AI compared.