Hedge-Bench
Agents analyze supplied company and investment documents for 102 open-ended financial research tasks. A model judge applies rubrics that check whether the reasoning covers the analytical steps identified by investment professionals.
Tasks run in Harbor containers. The README initially calls grading deterministic, but its scoring section explicitly uses semantic model judgments.
- Made by
- Trata
- Released
- June 2026
- Kind
- Finance
- Level
- Expert
- Scored
- Graded by a model
- Answers
- Agent task
- Tasks
- All public
- Measures
- Capability
- Subject
- Economics and business
- Language
- English
- Interaction
- Agentic
- Environment
- Terminal
- Items from
- Real work and use
- Refresh
- One fixed set
- Run by
- Anyone
- Occupations
- Business and financial operations
- In views
- Agents, Law and finance
- Strongest model tested
- Claude Fable 5.1 #8 on the capability index
- Last model added
- September 2026
- Scores refreshed
- November 20th, 2026
In the authors’ words
Modern AI agents are increasingly capable of routine financial-analysis tasks such as retrieving documents and updating spreadsheets. A harder challenge is forecasting an unknown future from incomplete information. Public markets provide a natural testbed: analysts routinely predict discrete events with information from that point in time. We present Hedge-Bench, a benchmark of 92 real-world environments derived from hedge fund analysts' reasoning traces paired with subsequently resolved prediction questions. Agents reason from frozen information snapshots, and their forecasts are evaluated against real-world outcomes. Across 11 frontier models and three trials per environment, joint success rates, which require both a correct forecast and over 60% research coverage, range from 2.54 to 31.52%. Using GRPO with turn-level credit assignment, we train Muse-Glimmer on 165 environments, increasing coverage on 22 held-out environments from 17.17% to 33.36% at low reasoning effort and from 30.19% to 47.02% at extra-high effort. The trained model at low effort also exceeds the base model's extra-high-effort coverage while using 47.9% fewer tokens. We also observe evidence of generalization: held-out forecasting accuracy at high effort improves by 10.06 percentage points, while out-of-distribution evaluation scores improve by 3.94 points on GDPval, 2.29 on GDP_pdf, 7.07 on Harvey LAB, and 4.96 on Frontier Finance under their respective metrics. Code: https://github.com/Trata-Inc/trata-hedge-bench-verified.
GPT-5.6 Sol leads with 31.5 among models released by September 2026
Best score per model from the benchmark's paper, read on November 20th, 2026. A paper’s table doesn’t add newer models. Where a model ran with several settings or harnesses, we keep its best.
The newest model with a score here came out in September 2026, and nobody has tested a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:
- GPT-6.2 Astra
- Claude Fable 5.2
-
Gemini 4 Carbon
- GPT-6 Astra
- and 12 more
- 1 GPT-5.6 Sol Hedge-Bench-Public; effort max 31.5
- 2 Claude Opus 5 Hedge-Bench-Public; effort max 29.0
- 3 Claude Fable 5.1 Hedge-Bench-Public; effort high 26.8
- 4 Grok 4.6 Hedge-Bench-Public; effort xhigh 26.4
- 5
Qwen 3.8 Max Hedge-Bench-Public; effort xhigh 25.7
- 6 GLM-5.3 Hedge-Bench-Public; effort max 14.1
- 7 GPT-5.6 Terra Hedge-Bench-Public; effort max 13.8
- 8
Muse Spark 1.3 Hedge-Bench-Public; effort xhigh 9.06
- 9 Kimi K3 Hedge-Bench-Public; effort max 8.70
- 10
Gemini 3.8 Flash Hedge-Bench-Public; effort high 4.71
- 11
DeepSeek V4 Pro 0813 Hedge-Bench-Public; effort max 2.54
Source: The Hedge-Bench paper
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | GPT-5.6 Sol | 31.5 | Hedge-Bench-Public; effort max | Jul 2026 | Official leaderboard |
| 2 | Claude Opus 5 | 29.0 | Hedge-Bench-Public; effort max | Jul 2026 | Official leaderboard |
| 3 | Claude Fable 5.1 | 26.8 | Hedge-Bench-Public; effort high | Sep 2026 | Official leaderboard |
| 4 | Grok 4.6 | 26.4 | Hedge-Bench-Public; effort xhigh | Aug 2026 | Official leaderboard |
| 5 | | 25.7 | Hedge-Bench-Public; effort xhigh | Aug 2026 | Official leaderboard |
| 6 | GLM-5.3 | 14.1 | Hedge-Bench-Public; effort max | Aug 2026 | Official leaderboard |
| 7 | GPT-5.6 Terra | 13.8 | Hedge-Bench-Public; effort max | Jul 2026 | Official leaderboard |
| 8 | | 9.06 | Hedge-Bench-Public; effort xhigh | Sep 2026 | Official leaderboard |
| 9 | Kimi K3 | 8.70 | Hedge-Bench-Public; effort max | Jul 2026 | Official leaderboard |
| 10 | | 4.71 | Hedge-Bench-Public; effort high | Sep 2026 | Official leaderboard |
| 11 | | 2.54 | Hedge-Bench-Public; effort max | Aug 2026 | Official leaderboard |
Number one over time
Lab claims count from the day the card came out, independent runs from the model’s release.
Source: The Hedge-Bench paper
Scores over time
All the scores we have, one dot each, so a model tested at several settings appears more than once.
Source: Official Hedge-Bench leaderboard
No lab model card reports it yet
We read the headline results of 540 launch posts from 57 labs, and none of them list this benchmark.
More in Law, finance and medicine
- MedXpertQA-MM In use Models answer expert-level medical exam questions that combine images with clinical records and test results. Scoring is accuracy.
Qwen 3.8 Max 80.4%
- HealthBench Professional In use 525 physician-authored tasks from real clinician chats (care consult, documentation, research) with physician-adjudicated rubrics. Claude Sonnet 5.5 69.2%
- HealthBench In use 5,000 realistic health conversations graded against physician-written rubrics. A length-adjusted variant penalizes verbosity. GPT-5 67.2%
- FinSearchComp In use Agents search the web to fetch live financial data, look up history, and run multi-step investigations, across 635 expert-written questions. Scored by answer accuracy.
Ling-3.1-flash 81.7%
- KorMedMCQA In use Models answer 7,469 multiple-choice questions from Korean licensing exams for doctors, nurses, pharmacists and dentists, held from 2012 to 2024. o1-preview 92.7%
- Harvey LAB-AA v1.1 In use Artificial Analysis's run of Harvey's Legal Agent Benchmark on Harvey's newer private set of 120 tasks across 24 practice areas. A deliverable with a material hallucination scores zero, and the headline is the share of judges who find every rubric criterion passed. GPT-6.2 Astra 15%
Data audit
All 9 checks pass. Every benchmark
Catalog record 2 pass
- Pass Task access All tasks are public.
- Pass Confidence Checked against a primary source.
Scores 3 pass 2 notes
- Pass Independent runs Scores from the benchmark’s paper (11 entries).
- Pass Models scored 11 models we track.
- Pass Newest score The newest is from September 2026, for Gemini 3.8 Flash.
- Note Frontier models Only 1 of the top 10 has a score. The strongest tested is Claude Fable 5.1 (#8).
- Note Headroom Scored as a plain number, with no fixed ceiling.
Leaderboard import 3 pass
- Pass Importer We read the board on arxiv.org: 11 rows for models we track at the last read.
- Pass Last read November 20th, 2026.
- Pass Read worked The last read worked.
Consistency 1 pass 1 note
- Note Lab claims No model has both a lab number and an independent run.
- Pass Known issues No published audit found problems.