ICLR 2026 Proprietary Agentic-System Cost Report

This repository contains a curated evidence set and report on the reported cost of proprietary agentic systems in ICLR 2026 accepted main-conference papers, with emphasis on papers that use Claude Code, Codex CLI, OpenAI Codex, or Codex-family models.

Main Outputs

• reports/iclr2026_proprietary_agentic_cost_report.pdf: final compiled PDF report.
• reports/iclr2026_proprietary_agentic_cost_report.tex: editable LaTeX source for the PDF.
• reports/iclr2026_proprietary_agentic_cost_report.md: markdown narrative version.
• reports/figures/: final figure assets used by the PDF.
• reports/source_index.md: curated paper-level evidence index.

Cleaned Repository Layout

• data/evidence_papers.tsv: the 16 retained evidence papers and their OpenReview PDF paths.
• data/agentic_candidate_papers.tsv: the broader 851-paper candidate list from the original scan.
• data/iclr2026_accepted.json: accepted-paper metadata used to define the full ICLR 2026 scope.
• texts/: compact local text conversions for the 16 retained evidence papers.
• results/tool_cost_hits.json: archived exact Claude Code/Codex and cost-term hit index from the original broad candidate scan.
• scripts/: small reproducibility scripts for fetching text, scanning hits, and regenerating figures.

The cleaned repository intentionally does not keep local PDFs. They were the largest redundant artifact, about 79 MB, and can be recovered from OpenReview using data/evidence_papers.tsv.

Rebuild The Figures

From the repository root:

python3 scripts/generate_report_figures.py

This rewrites the PDF and PNG files in reports/figures/.

Recompile The PDF

From the repository root:

cd reports
mkdir -p build
pdflatex -interaction=nonstopmode -halt-on-error -output-directory build iclr2026_proprietary_agentic_cost_report.tex
pdflatex -interaction=nonstopmode -halt-on-error -output-directory build iclr2026_proprietary_agentic_cost_report.tex
cp build/iclr2026_proprietary_agentic_cost_report.pdf iclr2026_proprietary_agentic_cost_report.pdf

The second pdflatex pass refreshes the table of contents and references.

Refresh Evidence Texts

The fetch script defaults to the curated 16-paper evidence manifest and deletes downloaded PDFs after converting them to text:

scripts/fetch_candidate_texts.sh

To keep PDFs locally during a rebuild:

KEEP_PDFS=1 scripts/fetch_candidate_texts.sh

To intentionally rerun the broad 851-paper candidate fetch, pass the broad manifest explicitly:

scripts/fetch_candidate_texts.sh data/agentic_candidate_papers.tsv

That broad run can create a large pdfs/ cache if KEEP_PDFS=1 is set.

Refresh The Hit Index

The checked-in results/tool_cost_hits.json was produced from the original broad candidate text set. The cleaned repository only keeps the 16 curated evidence texts, so running the scanner as-is will scan the retained evidence subset unless you first rebuild the broad text cache.

After updating whichever local text set you want to scan:

python3 scripts/search_tool_cost_hits.py

This rewrites results/tool_cost_hits.json.

Dependencies

• Python 3 with matplotlib and numpy for figure generation.
• pdflatex for PDF compilation.
• pdftotext and curl only if refreshing evidence texts from OpenReview.

Cleanup Policy

The repository keeps the final report, figure assets, curated evidence text, source metadata, and reproducibility scripts. It ignores temporary TeX products, rendered page previews, Python caches, and the large local PDF cache.