The Frontier AI Benchmark For Resume Extraction

ResumeBench evaluates 18 AI models and document parsers across 300 resumes, comparing accuracy, cost, and latency against human-verified ground truth.
Frontier agents measured on the benchmark

Why Resumes Are Hard

Measuring What AI Can Actually Extract
AI models can read resumes, but extracting every detail accurately is a different challenge.  ResumeBench tests how well leading AI systems handle different resume formats, from standard PDFs to complex layouts, scanned documents, and handwritten content. Every system is evaluated on the same dataset using consistent instructions and scoring criteria.
18
AI systems evaluated
300
Resumes tested
837
Pages analyzed
Benchmark at a Glance

What’s In The Benchmark

A Real-World Test of Resume Extraction
Real resume formats
Everyday PDFs, Canva and LaTeX designs, templates, handwritten scans, and rotated layouts. Roles span engineering, healthcare, data, and marketing.
Human-verified ground truth
Every resume has a reference record checked by people. Each field a system extracts is scored against it.
One test for all
Every system gets the same PDF, the same 14-section schema, and the same instructions.

The ResumeBench Leaderboard

See how leading AI models and document parsers compare on extraction accuracy, processing cost, and latency.
Model
F1 Score
Cost / 1K Pages
Latency
LlamaParse Extract Agentic Plus
87.16
$54
70s
Reducto Deep Extract
85.55
≈$40
65s
LlamaParse Extract Agentic
85.55
$23
55s
Claude Opus 4.8
85.51
$90
31s
Gemini 3.5 Flash
85.45
$35
43s
LlamaParse Extract Agentic Plus
F1 Score
87.16
Cost / 1K Pages
$54
Latency
70s
Reducto Deep Extract
F1 Score
85.55
Cost / 1K Pages
≈$40
Latency
65s
LlamaParse Extract Agentic
F1 Score
85.55
Cost / 1K Pages
$23
Latency
55s
Claude Opus 4.8
F1 Score
85.51
Cost / 1K Pages
$90
Latency
31s
Gemini 3.5 Flash
F1 Score
85.45
Cost / 1K Pages
$35
Latency
43s

How A Resume Is Scored

Scoring is exact and repeatable. There is no AI judge and no fuzzy matching.
Normalize
Standardize spaces, dashes, and quotation marks before comparison. Capitalization and punctuation still count.
Align
Match each extracted entry to its corresponding verified entry using identifying details, such as company, job title, and employment dates.
Compare
Check each field and bullet point for an exact match. A matching entry scores 1; a non-matching entry scores 0.
Score
Calculate precision, recall, and F1 for each section, average them for each resume, then average across all 300 resumes.

Careerflow Human data Labs

The team behind ResumeBench can test your model
ResumeBench is a Careerflow project. The resumes, the human-verified ground truth, and the evaluation methodology all come from our team.
We do the same work for AI teams who need to know how their model performs on their own data. Careerflow started as a career platform, which is why we know resumes so well. It is also where our network of 1M+ vetted experts comes from.
Custom eval sets
We collect and label test data that matches your documents, formats, and hard cases.
Expert ground truth
Trained reviewers from 100+ domains write & check the answers your scores depend on.
Scores and error reports
We score your model and show where it fails, field by field.

Evaluate Your Model Beyond The Benchmark

Public benchmarks measure performance on shared datasets. But real-world applications come with their own documents, edge cases, and quality requirements.
Scope
We agree on the task, the schema, and what a correct answer looks like.
Build the test set
We gather documents that match your real data, including the hard cases.
Label and review
Experts write the reference answers. A second pass checks each one
Score and report
You get scores and a field-level breakdown of where your model fails.

Have More Questions?

Here are some of the frequently asked questions from our customers
What is ResumeBench?

An open benchmark from Careerflow. It tests how well AI models and document parsers pull structured data out of resumes. It covers 18 systems, 300 resumes, and 837 pages.

Why not use an AI judge?

An AI judge can be wrong in ways that are hard to spot. Exact matching gives the same score every time, so anyone can check and repeat the results.

Can I see the data?

Yes. 20 resumes are published with their pages and reference answers. The full set of 300 is available from Careerflow on request.

Do you only build evals for resumes?

No. We build evals for documents, LLMs, agents, and multimodal models across 100+ domains.

How is my data handled?

Your data stays private. We use anonymized workflows and enterprise-grade security.

Get your submission
human-evaluated

Book a call and we'll review your agent's full trajectory set and return a leaderboard-ready result. After you book: a short scoping call to confirm the trajectory set and timing, then evaluation begins. No commitment on the first call.