Benchmark Almanac

Holdings-IndividualTaxBench

Claude Opus 5.5 76.3% leads since October 2026
Not in model cards

88 real client document packages turned into Form 1040 fields. Fact-level accuracy against practitioner-reviewed returns, with an LLM judge matching values.

Tasks come from Thrive's accounting firms and are not released. OpenAI holds equity in Thrive and the judge is an OpenAI model.

Made by
Thrive Holdings
Released
October 2026
Kind
Tasks from real jobs
Level
Expert
Scored
Checked automatically
Answers
Short answer
Tasks
Kept private
Strongest model tested
GPT-6 Astra #3 on the capability index
Last model added
September 2026
Scores refreshed
November 12th, 2026
13
Models scored on its own leaderboard

Claude Opus 5.5 leads with 76.3%

Best score per model from the benchmark's own leaderboard, read on November 12th, 2026. Where a model ran with several settings or harnesses, we keep its best. Every run is in the next chart.

13 models. Scroll sideways for the rest.

  1. 76.3 Claude Opus 5.5
  2. 75.5 Claude Fable 5.1
  3. 74.2 GPT-6.1 Sol
  4. 74.0 GPT-6 Astra
  5. 72.9 GPT-6 Sol
  6. 72.8 GLM-5.3
  7. 72.6 Kimi K3
  8. 72.3 Grok 4.7
  9. 71.5 Claude Sonnet 5
  10. 71.3 GLM-5.3-Flash
  11. 69.9 DeepSeek V4.1 Flash
  12. 67.6 GPT-6 Luna
  13. 64.4 DeepSeek V4 Pro 0813
  1. 1 Claude Opus 5.5 high 76.3%
  2. 2 Claude Fable 5.1 max 75.5%
  3. 3 GPT-6.1 Sol medium 74.2%
  4. 4 GPT-6 Astra medium 74%
  5. 5 GPT-6 Sol max 72.9%
  6. 6 GLM-5.3 max 72.8%
  7. 7 Kimi K3 high 72.6%
  8. 8 Grok 4.7 medium 72.3%
  9. 9 Claude Sonnet 5 xhigh 71.5%
  10. 10 GLM-5.3-Flash max 71.3%
  11. 11 DeepSeek V4.1 Flash max 69.9%
  12. 12 GPT-6 Luna medium 67.6%
  13. 13 DeepSeek V4 Pro 0813 xhigh 64.4%

Source: Official Holdings-IndividualTaxBench leaderboard

#ModelScoreSettingDateSource
1 Claude Opus 5.5 76.3% high Sep 2026 Official leaderboard
2 Claude Fable 5.1 75.5% max Sep 2026 Official leaderboard
3 GPT-6.1 Sol 74.2% medium Sep 2026 Official leaderboard
4 GPT-6 Astra 74% medium Sep 2026 Official leaderboard
5 GPT-6 Sol 72.9% max Sep 2026 Official leaderboard
6 GLM-5.3 72.8% max Aug 2026 Official leaderboard
7 Kimi K3 72.6% high Jul 2026 Official leaderboard
8 Grok 4.7 72.3% medium Sep 2026 Official leaderboard
9 Claude Sonnet 5 71.5% xhigh Jun 2026 Official leaderboard
10 GLM-5.3-Flash 71.3% max Aug 2026 Official leaderboard
11 DeepSeek V4.1 Flash 69.9% max Sep 2026 Official leaderboard
12 GPT-6 Luna 67.6% medium Sep 2026 Official leaderboard
13 DeepSeek V4 Pro 0813 64.4% xhigh Aug 2026 Official leaderboard

Performance, cost & time

Each point is one model at one reasoning effort: accuracy against the minutes it takes to prepare one return. The line joins the runs that no other run beats on both measures. We redrew this chart from the numbers on the official leaderboard.

50% 60% 70% 80% 05001,0001,5002,000 Time per return (s), lower is better Accuracy Claude Sonnet 5 (xhigh): Accuracy 71.5%, time per return 666 s Claude Sonnet 5 (max): Accuracy 71%, time per return 1,194 s Claude Sonnet 5 (high): Accuracy 67%, time per return 486 s Claude Sonnet 5 (medium): Accuracy 62.2%, time per return 312 s Kimi K3 (high): Accuracy 72.6%, time per return 1,020 s Kimi K3 (low): Accuracy 69.6%, time per return 516 s DeepSeek V4 Pro 0813 (xhigh): Accuracy 64.4%, time per return 576 s DeepSeek V4 Pro 0813 (high): Accuracy 61.7%, time per return 558 s GLM-5.3 (max): Accuracy 72.8%, time per return 738 s GLM-5.3 (high): Accuracy 72%, time per return 534 s GLM-5.3 (low): Accuracy 68.6%, time per return 306 s GLM-5.3-Flash (max): Accuracy 71.3%, time per return 1,656 s GLM-5.3-Flash (high): Accuracy 69.4%, time per return 798 s GLM-5.3-Flash (low): Accuracy 64%, time per return 1,026 s Claude Fable 5.1 (max): Accuracy 75.5%, time per return 906 s Claude Fable 5.1 (high): Accuracy 75.2%, time per return 456 s Claude Fable 5.1 (xhigh): Accuracy 75.2%, time per return 702 s Claude Fable 5.1 (medium): Accuracy 75.1%, time per return 366 s GPT-6 Astra (medium): Accuracy 74%, time per return 180 s GPT-6 Astra (max): Accuracy 73.5%, time per return 606 s GPT-6 Astra (high): Accuracy 73.1%, time per return 246 s GPT-6 Astra (xhigh): Accuracy 72.8%, time per return 306 s DeepSeek V4.1 Flash (max): Accuracy 69.9%, time per return 822 s DeepSeek V4.1 Flash (low): Accuracy 69.4%, time per return 276 s DeepSeek V4.1 Flash (high): Accuracy 69%, time per return 198 s Grok 4.7 (medium): Accuracy 72.3%, time per return 762 s Grok 4.7 (high): Accuracy 72%, time per return 948 s Grok 4.7 (xhigh): Accuracy 71.6%, time per return 1,164 s Grok 4.7 (low): Accuracy 67.6%, time per return 438 s Claude Opus 5.5 (xhigh): Accuracy 76.3%, time per return 504 s Claude Opus 5.5 (low): Accuracy 71.8%, time per return 228 s GPT-6 Luna (medium): Accuracy 67.6%, time per return 204 s GPT-6 Luna (high): Accuracy 64.6%, time per return 414 s GPT-6 Luna (xhigh): Accuracy 64%, time per return 438 s GPT-6 Luna (max): Accuracy 61.6%, time per return 924 s GPT-6 Sol (max): Accuracy 72.9%, time per return 378 s GPT-6 Sol (high): Accuracy 70.4%, time per return 156 s GPT-6 Sol (xhigh): Accuracy 70.4%, time per return 216 s GPT-6.1 Sol (high): Accuracy 73.6%, time per return 252 s GPT-6.1 Sol (xhigh): Accuracy 73.1%, time per return 294 s GPT-6.1 Sol (max): Accuracy 72.7%, time per return 744 s Claude Fable 5.1 (low): Accuracy 75.2%, time per return 312 s GPT-6 Astra (low): Accuracy 73.6%, time per return 144 s Claude Opus 5.5 (high): Accuracy 76.3%, time per return 366 s Claude Opus 5.5 (medium): Accuracy 75.7%, time per return 318 s GPT-6 Luna (low): Accuracy 54.9%, time per return 84 s GPT-6 Sol (medium): Accuracy 68.8%, time per return 108 s GPT-6 Sol (low): Accuracy 65%, time per return 96 s GPT-6.1 Sol (medium): Accuracy 74.2%, time per return 162 s GPT-6.1 Sol (low): Accuracy 69.9%, time per return 120 s Claude Opus 5.5 (high)Claude Opus 5.5 (medium)Claude Fable 5.1 (low)GPT-6.1 Sol (medium)GPT-6 Astra (low)GPT-6.1 Sol (low)GPT-6 Sol (medium)GPT-6 Sol (low)GPT-6 Luna (low) 50% 60% 70% 80% 05001,0001,5002,000 Time per return (s), lower is better Accuracy Claude Sonnet 5 (xhigh): Accuracy 71.5%, time per return 666 s Claude Sonnet 5 (max): Accuracy 71%, time per return 1,194 s Claude Sonnet 5 (high): Accuracy 67%, time per return 486 s Claude Sonnet 5 (medium): Accuracy 62.2%, time per return 312 s Kimi K3 (high): Accuracy 72.6%, time per return 1,020 s Kimi K3 (low): Accuracy 69.6%, time per return 516 s DeepSeek V4 Pro 0813 (xhigh): Accuracy 64.4%, time per return 576 s DeepSeek V4 Pro 0813 (high): Accuracy 61.7%, time per return 558 s GLM-5.3 (max): Accuracy 72.8%, time per return 738 s GLM-5.3 (high): Accuracy 72%, time per return 534 s GLM-5.3 (low): Accuracy 68.6%, time per return 306 s GLM-5.3-Flash (max): Accuracy 71.3%, time per return 1,656 s GLM-5.3-Flash (high): Accuracy 69.4%, time per return 798 s GLM-5.3-Flash (low): Accuracy 64%, time per return 1,026 s Claude Fable 5.1 (max): Accuracy 75.5%, time per return 906 s Claude Fable 5.1 (high): Accuracy 75.2%, time per return 456 s Claude Fable 5.1 (xhigh): Accuracy 75.2%, time per return 702 s Claude Fable 5.1 (medium): Accuracy 75.1%, time per return 366 s GPT-6 Astra (medium): Accuracy 74%, time per return 180 s GPT-6 Astra (max): Accuracy 73.5%, time per return 606 s GPT-6 Astra (high): Accuracy 73.1%, time per return 246 s GPT-6 Astra (xhigh): Accuracy 72.8%, time per return 306 s DeepSeek V4.1 Flash (max): Accuracy 69.9%, time per return 822 s DeepSeek V4.1 Flash (low): Accuracy 69.4%, time per return 276 s DeepSeek V4.1 Flash (high): Accuracy 69%, time per return 198 s Grok 4.7 (medium): Accuracy 72.3%, time per return 762 s Grok 4.7 (high): Accuracy 72%, time per return 948 s Grok 4.7 (xhigh): Accuracy 71.6%, time per return 1,164 s Grok 4.7 (low): Accuracy 67.6%, time per return 438 s Claude Opus 5.5 (xhigh): Accuracy 76.3%, time per return 504 s Claude Opus 5.5 (low): Accuracy 71.8%, time per return 228 s GPT-6 Luna (medium): Accuracy 67.6%, time per return 204 s GPT-6 Luna (high): Accuracy 64.6%, time per return 414 s GPT-6 Luna (xhigh): Accuracy 64%, time per return 438 s GPT-6 Luna (max): Accuracy 61.6%, time per return 924 s GPT-6 Sol (max): Accuracy 72.9%, time per return 378 s GPT-6 Sol (high): Accuracy 70.4%, time per return 156 s GPT-6 Sol (xhigh): Accuracy 70.4%, time per return 216 s GPT-6.1 Sol (high): Accuracy 73.6%, time per return 252 s GPT-6.1 Sol (xhigh): Accuracy 73.1%, time per return 294 s GPT-6.1 Sol (max): Accuracy 72.7%, time per return 744 s Claude Fable 5.1 (low): Accuracy 75.2%, time per return 312 s GPT-6 Astra (low): Accuracy 73.6%, time per return 144 s Claude Opus 5.5 (high): Accuracy 76.3%, time per return 366 s Claude Opus 5.5 (medium): Accuracy 75.7%, time per return 318 s GPT-6 Luna (low): Accuracy 54.9%, time per return 84 s GPT-6 Sol (medium): Accuracy 68.8%, time per return 108 s GPT-6 Sol (low): Accuracy 65%, time per return 96 s GPT-6.1 Sol (medium): Accuracy 74.2%, time per return 162 s GPT-6.1 Sol (low): Accuracy 69.9%, time per return 120 s Claude Opus 5.5 (high)Claude Opus 5.5 (medium)Claude Fable 5.1 (low)GPT-6.1 Sol (medium)GPT-6 Astra (low)GPT-6.1 Sol (low)GPT-6 Sol (medium)GPT-6 Sol (low)GPT-6 Luna (low)
ModelSettingAccuracyTime per return
Claude Opus 5.5 high 76.3% 366 s
Claude Opus 5.5 xhigh 76.3% 504 s
Claude Opus 5.5 medium 75.7% 318 s
Claude Fable 5.1 max 75.5% 906 s
Claude Fable 5.1 low 75.2% 312 s
Claude Fable 5.1 high 75.2% 456 s
Claude Fable 5.1 xhigh 75.2% 702 s
Claude Fable 5.1 medium 75.1% 366 s
GPT-6.1 Sol medium 74.2% 162 s
GPT-6 Astra medium 74% 180 s
GPT-6 Astra low 73.6% 144 s
GPT-6.1 Sol high 73.6% 252 s
GPT-6 Astra max 73.5% 606 s
GPT-6 Astra high 73.1% 246 s
GPT-6.1 Sol xhigh 73.1% 294 s
GPT-6 Sol max 72.9% 378 s
GPT-6 Astra xhigh 72.8% 306 s
GLM-5.3 max 72.8% 738 s
GPT-6.1 Sol max 72.7% 744 s
Kimi K3 high 72.6% 1,020 s
Grok 4.7 medium 72.3% 762 s
GLM-5.3 high 72% 534 s
Grok 4.7 high 72% 948 s
Claude Opus 5.5 low 71.8% 228 s
Grok 4.7 xhigh 71.6% 1,164 s
Claude Sonnet 5 xhigh 71.5% 666 s
GLM-5.3-Flash max 71.3% 1,656 s
Claude Sonnet 5 max 71% 1,194 s
GPT-6 Sol high 70.4% 156 s
GPT-6 Sol xhigh 70.4% 216 s
GPT-6.1 Sol low 69.9% 120 s
DeepSeek V4.1 Flash max 69.9% 822 s
Kimi K3 low 69.6% 516 s
DeepSeek V4.1 Flash low 69.4% 276 s
GLM-5.3-Flash high 69.4% 798 s
DeepSeek V4.1 Flash high 69% 198 s
GPT-6 Sol medium 68.8% 108 s
GLM-5.3 low 68.6% 306 s
GPT-6 Luna medium 67.6% 204 s
Grok 4.7 low 67.6% 438 s
Claude Sonnet 5 high 67% 486 s
GPT-6 Sol low 65% 96 s
GPT-6 Luna high 64.6% 414 s
DeepSeek V4 Pro 0813 xhigh 64.4% 576 s
GPT-6 Luna xhigh 64% 438 s
GLM-5.3-Flash low 64% 1,026 s
Claude Sonnet 5 medium 62.2% 312 s
DeepSeek V4 Pro 0813 high 61.7% 558 s
GPT-6 Luna max 61.6% 924 s
GPT-6 Luna low 54.9% 84 s

The best score, 76.3%, costs $5.00 a task

Runs on the cost frontier score higher than any run that's cheaper. GLM-5.3-Flash at max comes within 5 points of the best for $0.68 a task, 7 times cheaper. The costs are the leaderboard's own estimates, so compare them only within this benchmark.

Z.ai DeepSeek Anthropic OpenAI Cost frontier One model at each effort level
0% 20% 40% 60% 80% 100% $0.01 $0.10 $1.00 $10 $100 Cost per task, log scale Claude Opus 5.5, high GPT-6.1 Sol, medium DeepSeek V4.1 Flash, max GPT-6 Luna, low 0% 20% 40% 60% 80% 100% $0.01 $0.10 $1.00 $10 $100 Cost per task, log scale Claude Opus 5.5, high DeepSeek V4.1 Flash, max

Source: Official Holdings-IndividualTaxBench leaderboard

ModelEffortHarnessScoreCost per taskOutput tokens
Claude Opus 5.5 high 76.3% $5.00
Claude Opus 5.5 xhigh 76.3% $5.90
Claude Opus 5.5 medium 75.7% $4.90
Claude Fable 5.1 max 75.5% $16
Claude Fable 5.1 high 75.2% $11
Claude Fable 5.1 xhigh 75.2% $13
Claude Fable 5.1 low 75.2% $9.40
Claude Fable 5.1 medium 75.1% $10
GPT-6.1 Sol medium 74.2% $1.40
GPT-6 Astra medium 74% $8.90
GPT-6 Astra low 73.6% $7.90
GPT-6.1 Sol high 73.6% $1.50
GPT-6 Astra max 73.5% $13
GPT-6 Astra high 73.1% $10
GPT-6.1 Sol xhigh 73.1% $1.60
GPT-6 Sol max 72.9% $2.60
GLM-5.3 max 72.8% $3.60
GPT-6 Astra xhigh 72.8% $10
GPT-6.1 Sol max 72.7% $2.00
Kimi K3 high 72.6% $3.70
Grok 4.7 medium 72.3% $3.50
GLM-5.3 high 72% $3.10
Grok 4.7 high 72% $4.60
Claude Opus 5.5 low 71.8% $4.00
Grok 4.7 xhigh 71.6% $5.60
Claude Sonnet 5 xhigh 71.5% $3.20
GLM-5.3-Flash max 71.3% $0.68
Claude Sonnet 5 max 71% $4.00
GPT-6 Sol high 70.4% $1.80
GPT-6 Sol xhigh 70.4% $1.90
DeepSeek V4.1 Flash max 69.9% $0.42
GPT-6.1 Sol low 69.9% $1.30
Kimi K3 low 69.6% $3.20
GLM-5.3-Flash high 69.4% $0.53
DeepSeek V4.1 Flash low 69.4% $0.21
DeepSeek V4.1 Flash high 69% $0.20
GPT-6 Sol medium 68.8% $1.70
GLM-5.3 low 68.6% $2.20
Grok 4.7 low 67.6% $2.80
GPT-6 Luna medium 67.6% $0.11
Claude Sonnet 5 high 67% $2.70
GPT-6 Sol low 65% $1.50
GPT-6 Luna high 64.6% $0.14
DeepSeek V4 Pro 0813 xhigh 64.4% $1.00
GLM-5.3-Flash low 64% $0.47
GPT-6 Luna xhigh 64% $0.15
Claude Sonnet 5 medium 62.2% $2.40
DeepSeek V4 Pro 0813 high 61.7% $1.10
GPT-6 Luna max 61.6% $0.22
GPT-6 Luna low 54.9% $0.07

Number one over time

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

Anthropic
Own leaderboard 1 leader
Claude Opus 5.5 76.3%

Source: Official Holdings-IndividualTaxBench leaderboard

Scores over time

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

On the benchmark’s own leaderboard Best so far
0% 20% 40% 60% 80% 100% 2026 Near the ceiling (85%) Claude Opus 5.5 Kimi K3 0% 20% 40% 60% 80% 100% 2026 Near the ceiling (85%) Claude Opus 5.5

Source: Official Holdings-IndividualTaxBench leaderboard

No lab model card reports it yet

We read the headline results of 522 launch posts from 52 labs, and none of them list this benchmark.

More in Economic work

All 97 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
  • 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%
  • OfficeQA In use Grounded reasoning over a large corpus of U.S. Treasury Bulletin documents: locate tables and compute precise numeric answers. Claude Opus 4.7 86.3%