Benchmark Almanac

TreasuryQA Bench

Gemini 4 Hydrogen 82.4% leads since November 2026
Not in model cards

Agents search scanned Treasury volumes and read tables to find or calculate a number. A script checks whether the answer falls within the accepted range.

Veris ranks 12 models on 70 questions. The authors wrote 55 new questions and adapted 15 from OfficeQA Pro v2. They published three worked examples.

Made by
Veris AI and Reducto
Released
October 2026
Kind
Documents and spreadsheets
Level
School
Scored
Checked automatically
Answers
Agent task
Tasks
Public, plus a private test set
Measures
Capability
Language
English
In and out
Text, Documents, Tables → Actions, Text
Interaction
Agentic
Environment
APIs and tools
Refresh
One fixed set
Run by
A host
Based on
OfficeQA Pro V2
Strongest model tested
Gemini 4 Carbon #3 on the capability index
Last model added
November 2026
Scores refreshed
November 20th, 2026
12
Models scored on its own leaderboard

Gemini 4 Hydrogen leads with 82.4%

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.

  1. 1 Gemini 4 Hydrogen 82.4%
  2. 2 Gemini 4 Argon 81.9%
  3. 3 Gemini 4 Carbon 81%
  4. 4 Claude Opus 5.5 65.2%
  5. 5 Gemini 3.8 Flash 52.4%
  6. 6 GPT-6.1 Sol 52.4%
  7. 7 Claude Sonnet 5.5 42.9%
  8. 8 DeepSeek V4.1 Flash 41.9%
  9. 9 GLM-5.3-Flash 40%
  10. 10 Kimi K3 35.2%
  11. 11 Grok 4.7 22.9%
  12. 12 GPT-5.6 Terra 16.7%

Source: Official TreasuryQA Bench leaderboard

#ModelScore ±SettingDateSource
1 Gemini 4 Hydrogen 82.4% 74.3 to 90 Gemini CLI Nov 2026 Official leaderboard
2 Gemini 4 Argon 81.9% 73.8 to 89.5 Gemini CLI Sep 2026 Official leaderboard
3 Gemini 4 Carbon 81% 72.4 to 88.6 Gemini CLI Oct 2026 Official leaderboard
4 Claude Opus 5.5 65.2% 55.7 to 74.3 Claude Code Sep 2026 Official leaderboard
5 Gemini 3.8 Flash 52.4% ±10 Gemini CLI Sep 2026 Official leaderboard
6 GPT-6.1 Sol 52.4% ±10 Codex Sep 2026 Official leaderboard
7 Claude Sonnet 5.5 42.9% 33.3 to 52.9 Claude Code Sep 2026 Official leaderboard
8 DeepSeek V4.1 Flash 41.9% ±9.5 Pi Sep 2026 Official leaderboard
9 GLM-5.3-Flash 40% 31 to 50 Pi Aug 2026 Official leaderboard
10 Kimi K3 35.2% ±9.5 Pi Jul 2026 Official leaderboard
11 Grok 4.7 22.9% 16.2 to 30 Grok CLI Sep 2026 Official leaderboard
12 GPT-5.6 Terra 16.7% 10 to 23.8 Codex Jul 2026 Official leaderboard

Rankings by part

We read 2 parts from the benchmark’s own leaderboard: Clear and Very blurry. Each ranks models on one part of the test. Gemini 4 Argon leads Clear and Gemini 4 Hydrogen leads Very blurry. Arrows show how far a model moved from its overall rank among the models we track.

Clear

12 models we track have a score here.

  1. 1 Gemini 4 Argon ↑1 92%
  2. 2 Gemini 4 Hydrogen ↓1 87%
  3. 3 Gemini 4 Carbon 82%
  4. 4 GPT-6.1 Sol ↑2 74%
  5. 5 Claude Opus 5.5 ↓1 63%
  6. 6 Gemini 3.8 Flash ↓1 58%
  7. 7 Claude Sonnet 5.5 56%
  8. 8 GLM-5.3-Flash ↑1 55%
  9. 9 Kimi K3 ↑1 50%
  10. 10 DeepSeek V4.1 Flash ↓2 49%

The best score, 82.4%, costs $6.93 a task

Runs on the cost frontier score higher than any run that's cheaper. Gemini 4 Argon comes within 0.5 points of the best for $3.07 a task, about half the cost. The costs are the leaderboard's own estimates, so compare them only within this benchmark.

DeepSeek Anthropic OpenAI Google DeepMind Cost frontier One model at each effort level
0% 20% 40% 60% 80% 100% $0.10 $0.20 $0.50 $1.00 $2.00 $5.00 $10 Cost per task, log scale Gemini 4 Hydrogen Gemini 4 Argon Claude Opus 5.5 GPT-6.1 Sol DeepSeek V4.1 Flash 0% 20% 40% 60% 80% 100% $0.10 $1.00 $10 Cost per task, log scale Gemini 4 Hydrogen Claude Opus 5.5

Source: Official TreasuryQA Bench leaderboard

ModelEffortHarnessScoreCost per taskOutput tokens
Gemini 4 Hydrogen Gemini CLI 82.4% $6.93
Gemini 4 Argon Gemini CLI 81.9% $3.07
Gemini 4 Carbon Gemini CLI 81% $4.73
Claude Opus 5.5 Claude Code 65.2% $1.56
Gemini 3.8 Flash Gemini CLI 52.4% $1.75
GPT-6.1 Sol Codex 52.4% $0.60
Claude Sonnet 5.5 Claude Code 42.9% $1.21
DeepSeek V4.1 Flash Pi 41.9% $0.23
GLM-5.3-Flash Pi 40% $0.29
Kimi K3 Pi 35.2% $2.81
Grok 4.7 Grok CLI 22.9% $6.00
GPT-5.6 Terra Codex 16.7% $1.03

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

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

Google DeepMind
Own leaderboard 2 leaders

Source: Official TreasuryQA Bench 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% 2027 Near the ceiling (85%) Gemini 4 Hydrogen Gemini 4 Argon Claude Opus 5.5 Gemini 3.8 Flash GLM-5.3-Flash Kimi K3 GPT-5.6 Terra 0% 20% 40% 60% 80% 100% 2027 Near the ceiling (85%) Gemini 4 Hydrogen Claude Opus 5.5

Source: Official TreasuryQA 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 Economic work

All 210 in Economic work
  • JobBench In use 130 tasks from 35 occupations workers most want delegated. Agent works through messy reference files, graded by fact-anchored binary rubrics. GLM-5.3 61.4%
  • GDPval In use Real work tasks from 44 occupations with reference files. Experts blindly compare model deliverables with expert work. Metric is win-or-tie rate. Seed2.1 Pro 87.9%
  • GDPval-AA v2.1 In use Agents with shell and web access complete 220 GDPval gold tasks from 44 occupations. Blind pairwise comparisons give Elo ratings. Claude Opus 5.5 1,846
  • AA-Briefcase v1.1 In use Version 1.1 of Artificial Analysis's multi-hour office work evaluation, with Elo ratings fitted by Crowd-BT. Claude Opus 5.5 1,822
  • Agents' Last Exam v1.1 In use Agents complete 122 public long-horizon professional workflow tasks across 55 sub-industries. Deterministic graders check outcomes against hidden references. GPT-6.2 Astra 47.5%
  • CoWorkBench In use Agent completes long-horizon productivity tasks across computer science, finance, law and medicine. Qwen in-house scoring (method unpublished). Claude Fable 5 75.9%

Data audit

All 11 checks pass. Every benchmark

Catalog record 2 pass
  • Pass Task access Public tasks, plus a private test set.
  • Pass Confidence Checked against a primary source.
Scores 5 pass
  • Pass Independent runs Scores from the official leaderboard (12 entries).
  • Pass Models scored 12 models we track.
  • Pass Newest score The newest is from November 2026, for Gemini 4 Hydrogen.
  • Pass Frontier models 5 of the top 10 have a score, led by Gemini 4 Carbon (#3).
  • Pass Headroom The best is 82.4%, under 85%, so it still separates models.
Leaderboard import 3 pass
  • Pass Importer We read the board on veris.ai: 12 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