Benchmark Almanac

JFinBench

GPT-6 Astra 97.3% leads since November 2026
Not in model cards

Models answer questions from six Japanese certification exams in accounting, tax, financial planning, and banking. The authors report scores by exam.

Made by
Sakana AI
Released
November 2026
Kind
Finance
Level
University
Scored
Checked automatically
Answers
Multiple choice
Tasks
All public
Measures
Capability
Language
Japanese
In and out
Text → Text
Interaction
Single turn
Items from
Exams and contests
Refresh
One fixed set
Run by
Anyone
In views
Law and finance
Strongest model tested
GPT-6 Astra #4 on the capability index
Last model added
October 2026
Scores refreshed
November 20th, 2026
12
Models scored on its own leaderboard

GPT-6 Astra leads with 97.3%

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 GPT-6 Astra max, with tools 97.3%
  2. 2 Claude Fable 5 max, with tools 96.4%
  3. 3 Gemini 3.8 Flash max, with tools 94.2%
  4. 4 Kimi K3.1 Pro max, with tools 92.1%
  5. 5 DeepSeek V4.1 Flash 1017 max, with tools 88.6%
  6. 6 GLM-5.3-Flash max, with tools 87.6%
  7. 7 Kimi K3 max, with tools 86.1%
  8. 8 MiMo-V2.6 Pro thinking, with tools 85.3%
  9. 9 Qwen3.8 Flash default, with tools 83.3%
  10. 10 MiMo-V2.6 Flash thinking, with tools 73.2%
  11. 11 Kimi K2.6 thinking, with tools 70.4%
  12. 12 Mistral Large 4 thinking, with tools 70.1%

Source: Official JFinBench leaderboard

#ModelScoreSettingDateSource
1 GPT-6 Astra 97.3% max, with tools Sep 2026 Official leaderboard
2 Claude Fable 5 96.4% max, with tools Jun 2026 Official leaderboard
3 Gemini 3.8 Flash 94.2% max, with tools Sep 2026 Official leaderboard
4 Kimi K3.1 Pro 92.1% max, with tools Oct 2026 Official leaderboard
5 DeepSeek V4.1 Flash 1017 88.6% max, with tools Oct 2026 Official leaderboard
6 GLM-5.3-Flash 87.6% max, with tools Aug 2026 Official leaderboard
7 Kimi K3 86.1% max, with tools Jul 2026 Official leaderboard
8 MiMo-V2.6 Pro 85.3% thinking, with tools Sep 2026 Official leaderboard
9 Qwen3.8 Flash 83.3% default, with tools Aug 2026 Official leaderboard
10 MiMo-V2.6 Flash 73.2% thinking, with tools Sep 2026 Official leaderboard
11 Kimi K2.6 70.4% thinking, with tools Apr 2026 Official leaderboard
12 Mistral Large 4 70.1% thinking, with tools Oct 2026 Official leaderboard

Rankings by part

We read 6 parts from the benchmark’s own leaderboard, each ranking models on one part of the test, such as CPA, Tax accountant and FP. GPT-6 Astra leads 3 of them, and Claude Fable 5 leads the other 3. Arrows show how far a model moved from its overall rank among the models we track.

CPA

12 models we track have a score here.

  1. 1 GPT-6 Astra 99.6%
  2. 2 Claude Fable 5 98.6%
  3. 3 Gemini 3.8 Flash 97.8%
  4. 4 Kimi K3.1 Pro 96.1%
  5. 5 DeepSeek V4.1 Flash 1017 94.6%
  6. 6 GLM-5.3-Flash 93.5%
  7. 7 Kimi K3 91%
  8. 8 MiMo-V2.6 Pro 91%
  9. 9 Qwen3.8 Flash 89.6%
  10. 10 MiMo-V2.6 Flash 83.2%

The best score, 97.3%, costs $1.03 a task

Runs on the cost frontier score higher than any run that's cheaper. Claude Fable 5 at low comes within 4.7 points of the best for $0.06 a task, 17 times cheaper. The costs are the leaderboard's own estimates, so compare them only within this benchmark.

Anthropic Z.ai Google DeepMind OpenAI DeepSeek Cost frontier One model at each effort level
0% 20% 40% 60% 80% 100% $0.001 $0.01 $0.10 $1.00 $10 Cost per task, log scale GPT-6 Astra, max, with tools Claude Fable 5, low Gemini 3.8 Flash, low DeepSeek V4.1 Flash 1017, low GLM-5.3-Flash, low 0% 20% 40% 60% 80% 100% $0.001 $0.01 $0.10 $1.00 $10 Cost per task, log scale GPT-6 Astra, max, with tools Gemini 3.8 Flash, max

Source: Official JFinBench leaderboard

ModelEffortHarnessScoreCost per taskOutput tokens
GPT-6 Astra max, with tools 97.3% $1.03 13,780
Claude Fable 5 max, with tools 96.4% $2.26 36,300
Claude Fable 5 max 95.8% $0.56 10,854
Gemini 3.8 Flash max, with tools 94.2% $0.38 28,200
GPT-6 Astra max 93.6% $0.25 4,751
Claude Fable 5 low 92.6% $0.06 938
GPT-6 Astra low 92.1% $0.04 682
Kimi K3.1 Pro max, with tools 92.1% $0.49 22,540
Gemini 3.8 Flash max 88.8% $0.03 6,567
DeepSeek V4.1 Flash 1017 max, with tools 88.6% $0.05 133,020
GLM-5.3-Flash max, with tools 87.6% $0.22 132,080
Kimi K3 max, with tools 86.1% $0.34 15,310
MiMo-V2.6 Pro thinking, with tools 85.3% $0.09 82,180
Gemini 3.8 Flash low 85.1% $0.005 1,146
Kimi K3.1 Pro max 83.4% $0.15 14,262
Qwen3.8 Flash default, with tools 83.3% $0.09 136,760
Kimi K3 max 79.9% $0.25 16,636
Kimi K3.1 Pro low 78.4% $0.04 3,928
MiMo-V2.6 Flash thinking, with tools 73.2% $0.04 106,570
DeepSeek V4.1 Flash 1017 max 73.1% $0.007 23,495
Kimi K3 low 72.5% $0.03 1,794
DeepSeek V4.1 Flash 1017 low 71.7% $0.003 8,953
Kimi K2.6 thinking, with tools 70.4% $0.35 50,580
Mistral Large 4 thinking, with tools 70.1% $0.36 83,020
Qwen3.8 Flash default 62.5% $0.02 43,206
MiMo-V2.6 Pro thinking 62.1% $0.03 33,863
Kimi K2.6 thinking 59.1% $0.06 14,162
Mistral Large 4 thinking 58.2% $0.06 29,044
GLM-5.3-Flash max 57.7% $0.01 26,430
MiMo-V2.6 Flash thinking 50.9% $0.01 34,421
GLM-5.3-Flash low 50.8% $0.002 3,274

Number one over time

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

OpenAI
Own leaderboard Passed 85% in Nov 2026, by GPT-6 Astra
GPT-6 Astra 97.3%

Source: Official JFinBench leaderboard

Claude Fable 5 on the official leaderboard passed 85% in June 2026

Models released before it came out already scored that high. 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 2027 Near the ceiling (85%) GPT-6 Astra Claude Fable 5 Kimi K2.6 0% 20% 40% 60% 80% 100% 2026 2027 Near the ceiling (85%) GPT-6 Astra Kimi K2.6

Source: Official JFinBench 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

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 official leaderboard (12 entries).
  • Pass Models scored 12 models we track.
  • Pass Newest score The newest is from October 2026, for Kimi K3.1 Pro.
  • Note Frontier models 1 of the top 10 has a score, which is fine for a retired test. The strongest tested is GPT-6 Astra (#4).
  • Note Headroom Saturated. The best is 97.3%, past 85% since June 2026.
Leaderboard import 3 pass
  • Pass Importer We read the board on pub.sakana.ai: 31 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