Holdings-IndividualTaxBench
Claude Opus 5.5 76.3% leads since Oct 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 Oct 2026 Tasks from real jobs Tasks kept private
Last model added Sep 2026
Scores refreshed 2 Nov 2026
- Level
- Expert Specialists with years of training, often over hours of work
- Scored
- Checked automatically A program checks the result against an answer key, unit tests or the expected final state.
- Answers
- Short answer A number, word or short string matched to a key.
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 2 Nov 2026. Where a model ran with several settings or harnesses, we keep its best. Every run is in the next chart.
- 1 Claude Opus 5.5 high 76.3%
- 2 Claude Fable 5.1 max 75.5%
- 3 GPT-6.1 Sol medium 74.2%
- 4 GPT-6 Astra medium 74%
- 5 GPT-6 Sol max 72.9%
- 6 GLM-5.3 max 72.8%
- 7 Kimi K3 high 72.6%
- 8 Grok 4.7 medium 72.3%
- 9 Claude Sonnet 5 xhigh 71.5%
- 10 GLM-5.3-Flash max 71.3%
- 11
DeepSeek V4.1 Flash max 69.9% - 12 GPT-6 Luna medium 67.6%
- 13 DeepSeek V4 Pro 0813 xhigh 64.4%
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 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 |
The best score, 76.3%, costs $5.00 a task
Runs on the cost frontier score higher than any run that's cheaper. GPT-6.1 Sol at medium comes within 2.1 points of the best for $1.40 a task, 4 times cheaper. The costs are the leaderboard's own estimates, so compare them only within this benchmark.
DeepSeek Anthropic OpenAI
Cost frontier
One model at each effort level
| Model | Effort | Harness | Score | Cost per task | Output 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% | $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% | $42 | ||
| GPT-6.1 Sol | low | 69.9% | $1.30 | ||
| Kimi K3 | low | 69.6% | $3.20 | ||
| GLM-5.3-Flash | high | 69.4% | $53 | ||
| DeepSeek V4.1 Flash | low | 69.4% | $22 | ||
| DeepSeek V4.1 Flash | high | 69% | $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% | $11 | ||
| Claude Sonnet 5 | high | 67% | $2.70 | ||
| GPT-6 Sol | low | 65% | $1.50 | ||
| GPT-6 Luna | high | 64.6% | $14 | ||
| DeepSeek V4 Pro 0813 | xhigh | 64.4% | $1.00 | ||
| GLM-5.3-Flash | low | 64% | $47 | ||
| GPT-6 Luna | xhigh | 64% | $15 | ||
| Claude Sonnet 5 | medium | 62.2% | $2.40 | ||
| DeepSeek V4 Pro 0813 | high | 61.7% | $1.10 | ||
| GPT-6 Luna | max | 61.6% | $22 | ||
| GPT-6 Luna | low | 54.9% | $7.20 |
Number one over time
Lab claims count from the day the card came out, independent runs from the model’s release.
Anthropic
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
No lab model card reports it yet
We read the headline results of 479 launch posts from 51 labs, and none of them list this benchmark.
More 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).
Qwen3.8-Omni-Flash 75.3% - 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%