Science's Last Exam: AItonomy's agent research bench

The AItonomy Foundation launched Science’s Last Exam (SLE), a benchmark aimed at whether AI agents can do scientific research work, not just answer exam questions. The public set lists 113 tasks across nine disciplines, with six scored capability axes plus an end-to-end research assessment. Wet-lab execution is marked In construction and unscored; end-to-end research is marked To be expanded.

Leaderboard snapshot

On the publisher board (published 4 November 2026; figures synced on the HF card as of 5 Nov):

• Best composite so far: Claude Opus 5 with Claude Code at High effort, 37.50 / 100.
• GPT-6.1 Sol follows at 36.15.
• End-to-end research scores are 0 for all currently listed models (a published result, not a fixed rule that zeros are permanent).
• Claude Opus 5.5 and GPT-6 Astra are N/A: the site says provider-enforced biological safeguards can restrict biology research workflows; results wait on research access.

Attribute every number to AItonomy’s leaderboard. Harness, effort and cost columns are the publisher’s.

What you can download today

Do not claim the whole project is open-sourced.

• GitHub Gen-Verse/Science-Last-Exam is public (README cites Apache-2.0 for software and CC-BY-4.0 for data).
• As of triage ~09:31 IST on 6 November 2026, Hugging Face marked the task-card and data datasets as not gated, and an unauthenticated fetch of task_cards.parquet succeeded. The reference dataset remains manual-gated.
• The Hugging Face dataset card text may still say “Release status: private preview” / permission required. Treat that wording as stale relative to the download behaviour checked at triage. Hidden reference assets stay outside the solver’s environment on purpose.

Why this matters

If you ship science or agent stacks, SLE is a rare end-to-end-shaped bench where today’s best listed composites still sit in the mid-30s and nobody on the board clears the end-to-end axis. That is a useful brake on “models can already do research” marketing, including around life-science products such as GPT-6 Rosalind or narrower science tooling such as Claude’s UV sky-map work. Builders who want to reproduce should start from the public GitHub harness plus the public task cards and data, then apply for reference access rather than assuming every asset is free to pull.