Benchmark Almanac

Policy Agent

GPT-5.4 100% leads since March 2026
Not in model cards

Models approve expenses, reject them, or request human review. Ramp compares their decisions with human expense reviews.

Made by
Ramp Labs
Released
March 2026
Kind
Finance
Level
University
Scored
Checked automatically
Answers
Multiple choice
Tasks
Kept private
Measures
Capability
Language
English
In and out
Text → Text
Interaction
Single turn
Items from
Real work and use
Refresh
One fixed set
Run by
A host
In views
Law and finance
Strongest model tested
GPT-5.4 #28 on the capability index
Last model added
March 2026
Scores refreshed
November 20th, 2026
10
Models scored on its own leaderboard

GPT-5.4 leads with 100% among models released by March 2026

Best score per model from the benchmark's own leaderboard, read on November 20th, 2026. Where a model ran with several settings or harnesses, we keep its best. Every run is in the next chart.

The newest model with a score here came out in March 2026, and nobody has tested a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:

  1. 1 GPT-5.4 low 100%
  2. 2 Claude Sonnet 4.6 none 114.7%
  3. 3 GPT-5.2 low 118.3%
  4. 4 Gemini 3 Pro low 151.4%
  5. 5 Gemini 3 Flash medium 162.1%
  6. 6 Gemini 3.1 Pro high 162.1%
  7. 7 Kimi K2.5 163.9%
  8. 8 Claude Opus 4.6 high 164.4%
  9. 9 GPT-5 mini high 173.9%
  10. 10 GPT-5.1 low 218.8%

Source: Official Policy Agent leaderboard

#ModelScoreSettingDateSource
1 GPT-5.4 100% low Mar 2026 Official leaderboard
2 Claude Sonnet 4.6 114.7% none Feb 2026 Official leaderboard
3 GPT-5.2 118.3% low Dec 2025 Official leaderboard
4 Gemini 3 Pro 151.4% low Nov 2025 Official leaderboard
5 Gemini 3 Flash 162.1% medium Dec 2025 Official leaderboard
6 Gemini 3.1 Pro 162.1% high Feb 2026 Official leaderboard
7 Kimi K2.5 163.9% Jan 2026 Official leaderboard
8 Claude Opus 4.6 164.4% high Feb 2026 Official leaderboard
9 GPT-5 mini 173.9% high Aug 2025 Official leaderboard
10 GPT-5.1 218.8% low Nov 2025 Official leaderboard

More effort adds a median 2 points

Models the leaderboard ran at two or more effort levels. We don’t have enough cost data here to say what the extra effort costs.

Lowest effort to highest, light to dark
GPT-5.1
none low medium high
274.4%
GPT-5 mini
minimal low medium high
267.3%
Claude Opus 4.6
none
166.8%
Gemini 3 Flash
minimal low medium
205.8%
Gemini 3.1 Pro
low
164.4%
Gemini 3 Pro
low high
159.6%
GPT-5.2
none low medium high
140.7%
Claude Sonnet 4.6
none low high
162.1%
GPT-5.4
none low medium
116.9%

Source: Official Policy Agent leaderboard

ModelEffortHarnessScoreCost per taskOutput tokens
GPT-5.1 none 274.4%
GPT-5.1 high 272%
GPT-5 mini minimal 267.3%
GPT-5.1 medium 247.2%
GPT-5.1 low 218.8%
Gemini 3 Flash low 205.8%
GPT-5 mini low 194%
Gemini 3 Flash minimal 190.4%
GPT-5 mini medium 188.1%
GPT-5 mini high 173.9%
Claude Opus 4.6 low 166.8%
Gemini 3 Flash high 165.6%
Claude Opus 4.6 medium 165.6%
Claude Opus 4.6 none 165.6%
Claude Opus 4.6 high 164.4%
Gemini 3.1 Pro low 164.4%
Kimi K2.5 163.9%
Gemini 3 Flash medium 162.1%
Claude Sonnet 4.6 low 162.1%
Gemini 3.1 Pro high 162.1%
Claude Sonnet 4.6 medium 160.9%
Gemini 3 Pro high 159.6%
Claude Sonnet 4.6 high 157.3%
Gemini 3 Pro low 151.4%
GPT-5.2 none 140.7%
GPT-5.2 high 134.8%
GPT-5.2 medium 124.1%
GPT-5.2 low 118.3%
GPT-5.4 none 116.9%
Claude Sonnet 4.6 none 114.7%
GPT-5.4 medium 110.6%
GPT-5.4 high 107.2%
GPT-5.4 low 100%

2 models have held number one on the official leaderboard since its release in March 2026

Lab claims count from the day the card came out, independent runs from the model’s release.

Anthropic OpenAI
Own leaderboard Not tested since Mar 2026
GPT-5.4 100%

Source: Official Policy Agent leaderboard

Scores over time

All the scores we have, one dot each, so a model tested at several settings appears more than once.

Plot with other benchmarks
On the benchmark’s own leaderboard Best so far
0% 20% 40% 60% 80% 100% 2026 GPT-5.4 GPT-5.2 GPT-5 mini 0% 20% 40% 60% 80% 100% 2026 GPT-5.4 GPT-5 mini

Source: Official Policy Agent 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

All 637 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

8 of 9 checks pass, 1 warning. Look first at the Newest score check. The newest came 8 months ago, in March 2026, for GPT-5.4. Every benchmark

Catalog record 2 pass
  • Pass Task access The tasks are private.
  • Pass Confidence Checked against a primary source.
Scores 1 warning 2 pass 2 notes
  • Pass Independent runs Scores from the official leaderboard (10 entries).
  • Pass Models scored 10 models we track.
  • Warning Newest score The newest came 8 months ago, in March 2026, for GPT-5.4.
  • Note Frontier models None of the top 10 has a score. The strongest tested is GPT-5.4 (#28).
  • Note Headroom Lower is better, so there’s no ceiling to reach.
Leaderboard import 3 pass
  • Pass Importer We read the board on builders.ramp.com: 33 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.

How we check