A page-by-page, rubric-guided multimodal LLM workflow graded an authentic handwritten exam in about three hours at roughly $100 in token costs, versus about $3,500 in TA time¶

Claim · All claims
Evidence · 1 study (2 entries) · design r2 · q2
Currency · newest study 2026 · checked against its sources 2026-10-20 (automated, 2 of 2 entries) · not reviewed by a person

Subclaims¶

Reading the evidence codes

q is the design tier, i the impact, n the sample, and the kind with its r the rigour judged within that kind: what each code means.

This claim: no standardized effect is recorded in observations/ for it yet, so no benchmark reading is possible.

q2 i? Each page submission took between 50 and 60 seconds; running up to 30 agents in parallel allowed grading the exams in about three hours. → Jan Cvengros and Gerd Kortemeyer 2026 q2 i? The workflow used 32.15 million tokens costing approximately $100, compared to about $3,500 for TA grading of the whole exam for all 459 students. → Jan Cvengros and Gerd Kortemeyer 2026 (2)

Evidence¶

Jan Cvengros and Gerd Kortemeyer 2026¶

Jan Cvengros and Gerd Kortemeyer. (2026). Assisting the grading of a handwritten general chemistry exam with artificial intelligence. Discover Artificial Intelligence. https://doi.org/10.1007/s44163-026-01606-4

q2 · i? · design · r2

Operational workflow report (Sect. 3.2) mirroring an intended production deployment: page images of student work plus rubric pages were submitted one page at a time to gpt-o4-mini/high-vision via ephemeral URLs, producing structured JSON output per part.

"Each submission took between 50 and 60 s. At times we ran 30 agents in parallel, allowing us to grade the exams in about three hours"

Jan Cvengros and Gerd Kortemeyer 2026 (2)¶

q2 · i? · design · r2

Cost accounting in Sect. 3.2 compares the token cost with 16 TAs × 5 h × $44/h ≈ $3,500 for grading the whole exam for all 459 students, using the ETH-stipulated gross hourly income of CHF 30.70 plus employer contributions.

"We used 32.15 million tokens, costing approximately $100 in total"

Discussion¶

In a real academic pilot, AISSA processed 90 presentations reliably at 1–3 minutes per submission and an estimated cost of $0.06–0.07 USD per evaluation — related
Human scoring and feedback for HDR takes about 11 minutes 30 seconds per learner, while GPT-based scoring and feedback is nearly instant — related
Structure and layout of the exam pages strongly contribute to the success of the multimodal grading method and, for now, appear to be required — related
An on-premises LLM autograder graded a 200-student course at 1–3 minutes wall-clock time per submission across a full term — related
AI grading of the exam is highly stable across five independent runs at the total-score level (ICC(A,1) = 0.967), with lower but still strong cell-level stability (ICC(A,1) = 0.836) — related
AI-assigned total scores on a handwritten chemistry exam align strongly with TA-assigned totals (R2 = 0.91), while per-problem agreement is lower (R2 = 0.61–0.85) — related
Early adopters report time savings from ChatGPT automation, including a Wharton professor's exam-writing productivity increase of 100% — related
LLM-based extraction of three variables from 156 meta-analyses cost about $50 in computing fees, far less than estimated human-assistant labor costs — related

Page metadata

Field	Value
`type`	`claim`
`id`	`ai-grading-workflow-feasibility-cost`
`title`	`A page-by-page, rubric-guided multimodal LLM workflow graded an authentic handwritten exam in about three hours at roughly $100 in token costs, versus about $3,500 in TA time`
`description`	`A page-by-page, rubric-guided multimodal LLM workflow graded an authentic handwritten exam in about three hours at roughly $100 in token costs, versus about $3,500 in TA time`
`status`	`review`
`evidence_strength`	`moderate`
`newest_evidence`	`2026`
`source_checked`	`2026-10-20`
`generated`	by `process:wiki-ingest` on `2026-10-10`
`trust`	unverified — no one has confirmed this page's content
`sources`	2 — `jan-cvengros-and-gerd-kortemeyer-2026`, `jan-cvengros-and-gerd-kortemeyer-2026-2`

Frontmatter is the page's machine-readable record — see the Schema & Guide.