Skip to content

OpenAI · Vision LLM

GPT-5.6 on ResumeBench: 84.7 F1, rank 6 of 18

GPT-5.6 ranks 6th of 18 on ResumeBench: 84.7 F1, $74 per 1,000 pages, 38 seconds per resume. It is 2.4 points behind LlamaParse Extract Agentic Plus. LlamaParse Extract Agentic Plus scores 87.2 for $54.

3rd of 11 vision LLMs. Strongest on content as images (+4.8 points), distractor layouts (+4.7 points) and wording traps (+4.6 points). Weakest on scanned & rotated (−11.4 points), rotated & warped pages (−9.7 points) and hidden text & prompt injections (−7.0 points).

Rank
6/18
F1
84.7
Precision
85.9
Recall
87.2
Perfect resumes
36/300
Per 1,000 pages
$74
Median per resume
38 s

Averaged over 300 resumes. Dates (scored separately): 97.6. 90% finish within 66 s.

Where GPT-5.6 wins and loses

F1 by kind of resume, against the median system.

Reading the page

GPT-5.6 by reading the page
Kind of resume# of resumesF1MedianGapRank
Handwritten953.251.5+1.86th
Image only2865.471.4−5.913th
Sidebar or grid8281.781.9−0.311th
Photo4585.584.0+1.56th

Distractors

GPT-5.6 by distractors
Kind of resume# of resumesF1MedianGapRank
Rotated & warped pages2173.282.9−9.713th
Hidden text & prompt injections1477.584.5−7.013th
Content as images893.088.2+4.82nd
Wording traps1689.184.5+4.61st
Typos kept on purpose3780.683.7−3.112th

Extraction challenges

GPT-5.6 by extraction challenges
Kind of resume# of resumesF1MedianGapRank
Many entries8384.581.9+2.63rd
Long bullet lists10184.981.4+3.52nd
Split across pages8584.982.2+2.63rd
Big skills list6085.283.6+1.56th
Custom sections10582.280.8+1.36th
Combined headings3784.182.7+1.44th
Promotions8885.083.5+1.57th
Partial dates13885.783.4+2.22nd
Overlapping jobs6983.281.7+1.47th
Accented text1386.585.5+1.07th

Noise

GPT-5.6 by noise
Kind of resume# of resumesF1MedianGapRank
Repeated header6786.885.2+1.74th
References5587.985.2+2.75th
Busy page1884.080.1+3.91st

Source

GPT-5.6 by source
Kind of resume# of resumesF1MedianGapRank
Everyday resumes21286.885.1+1.74th
Expert-built (Canva / LaTeX)3285.283.6+1.63rd
Resume-builder templates1083.781.2+2.57th
Scanned & rotated1971.682.9−11.413th
Distractor layouts1890.185.3+4.71st

Layout

GPT-5.6 by layout
Kind of resume# of resumesF1MedianGapRank
Single column19585.584.0+1.63rd
Two-column2389.086.8+2.15th
Two-column sidebar5478.380.0−1.611th
Three-column2085.684.8+0.76th
Table grid894.594.1+0.58th

Length

GPT-5.6 by length
Kind of resume# of resumesF1MedianGapRank
Short14887.287.1+0.09th
Medium9085.684.7+1.04th
Long6277.774.6+3.05th

By section of the resume

F1 by resume section, against the median system.

GPT-5.6 by resume section
SectionResumesF1MedianGap
Contact details30094.193.5+0.5
Experience30088.486.3+2.1
Education29791.989.4+2.5
Summary29083.481.7+1.7
Certifications25577.981.1−3.2
Awards9581.275.4+5.8
Skills28793.292.1+1.1
Languages16885.984.3+1.5
Custom sections21251.249.9+1.3
Interests6491.491.7−0.4
Volunteering7785.784.5+1.1
Profiles3769.466.7+2.7
Projects8088.283.2+5.0
Publications5177.878.6−0.8

On the published resumes

Best and worst of the 20 public resumes.

GPT-5.6 on published resumes
ResumeWhat makes it hardF1
Head of Developer Experience (Internal title: Jedi Master of Analytics) Unusual job titles93.6
Creative DirectorAccented text in a styled CV90.8
Head of Events & Public Space OperationsTable grid with emoji bullets90.0
Senior Full-Stack EngineerLaTeX CV with a running header69.7
senior Consultant - package ImplementationHandwritten resume47.2
Senior ocularist and practice ownerHandwritten resume, six pages40.6

How it was run

OpenAI API, model gpt-5.6-sol, page images in, JSON schema out. Same PDF, schema and instructions as every other system, one run per resume (4 November 2026).

GPT-5.6 on OpenAI’s site

Run the open-source harness: list its pipelines, or grade your own predictions (GitHub):

$ pip install "resume-bench[runners] @ git+https://github.com/careerflow/ResumeBench"
$ hf auth login   # the dataset is gated: accept its terms on Hugging Face first
$ resume-bench download
$ resume-bench providers
$ resume-bench grade-file my_predictions.jsonl --split test

Closest alternatives

Nearest in F1.

Embed this result

Shows its current rank. Updates with each run.

GPT-5.6: #6 on ResumeBench resume extraction benchmark
<a href="https://www.resumebench.ai/systems/gpt-5-6"><img src="https://www.resumebench.ai/badge/gpt-5-6.svg" alt="GPT-5.6: #6 on ResumeBench resume extraction benchmark" height="28"></a>

Choosing a resume extraction model

All 18 models · Leaderboard

Frequently asked questions (FAQs)

How accurate is GPT-5.6 at resume extraction?

84.7 F1 on ResumeBench's 300 resumes, 6th of 18 models.

How much does GPT-5.6 cost to parse resumes?

$74 per 1,000 pages, about $0.207 per resume.

How fast is GPT-5.6?

38 seconds per resume (median).

Which resumes is GPT-5.6 best at?

Content as images, distractor layouts and wording traps. Weakest on scanned & rotated, rotated & warped pages and hidden text & prompt injections.

What is a cheaper alternative to GPT-5.6?

LlamaParse Extract Agentic Plus: 87.2 F1 for $54 per 1,000 pages.