MortgageTax & TaxEval v2

Visual tax-certificate extraction (MortgageTax) and tax regulation QA (TaxEval v2).26 Oct 2026

One row per model and compatible score measure. Mean = equal-weight arithmetic average of imported reports. Distinct configurations reported for the same model in a source table remain separate. Reports can overlap; the range is not a confidence interval. Unbundle to inspect each observation.
		Evidence	What the result shows
Muse Spark →Meta	77.68% 1 report	MortgageTax & TaxEval v2[2]Cost In / Out: N/A; Latency: 57.52s	Leads this displayed comparison.
Grok 3 →xAI	75.88% 1 report	MortgageTax & TaxEval v2[2]Cost In / Out: $3 / $15; Latency: 13.06s	#2 of 166 in this displayed comparison; 1.8 pp below the leader.
GLM 5.3 Flash →Z.ai	75.59% 1 report	MortgageTax & TaxEval v2[2]Cost In / Out: $0.075 / $0.25; Latency: 53.06s	#3 of 166 in this displayed comparison; 2.09 pp below the leader.
Qwen 3.7 Max →Qwen	75.31% 1 report	MortgageTax & TaxEval v2[2]Cost In / Out: $2.5 / $7.5; Latency: 57.68s	#4 of 166 in this displayed comparison; 2.37 pp below the leader.
Claude Opus 4.5 →AnthropicCost In / Out: $5 / $25; Latency: 47.45s	74.86% 1 report	MortgageTax & TaxEval v2[2]Cost In / Out: $5 / $25; Latency: 47.45s	#5 of 166 in this displayed comparison; 2.82 pp below the leader.
Claude Opus 4.5 →AnthropicCost In / Out: $5 / $25; Latency: 14.88s	74.33% 1 report	MortgageTax & TaxEval v2[2]Cost In / Out: $5 / $25; Latency: 14.88s	#6 of 166 in this displayed comparison; 3.35 pp below the leader.
o1 →OpenAI	74.28% 1 report	MortgageTax & TaxEval v2[2]Cost In / Out: $15 / $60; Latency: 19.48s	#7 of 166 in this displayed comparison; 3.4 pp below the leader.
Claude 3.7 Sonnet →AnthropicCost In / Out: $3 / $15; Latency: 43.31s	74.04% 1 report	MortgageTax & TaxEval v2[2]Cost In / Out: $3 / $15; Latency: 43.31s	#8 of 166 in this displayed comparison; 3.64 pp below the leader.
Qwen 3 Max Preview →Qwen	73.96% 1 report	MortgageTax & TaxEval v2[2]Cost In / Out: $1.2 / $6; Latency: 21.41s	#9 of 166 in this displayed comparison; 3.72 pp below the leader.
MiMo V2.5 Pro →Xiaomi	73.79% 1 report	MortgageTax & TaxEval v2[2]Cost In / Out: $0.435 / $0.87; Latency: 62.83s	#10 of 166 in this displayed comparison; 3.89 pp below the leader.
Claude Opus 4.1 →AnthropicCost In / Out: $15 / $75; Latency: 26.64s	73.67% 1 report	MortgageTax & TaxEval v2[2]Cost In / Out: $15 / $75; Latency: 26.64s	#11 of 166 in this displayed comparison; 4.01 pp below the leader.
Claude Opus 5 →Anthropic	73.6% 2 reports · range 72.06–75.14%	MortgageTax & TaxEval v2[2][1] View 2 reports 75.14 % · TaxEval · Accuracy · Cost In / Out: $5 / $25; Latency: 58.34s [2] 72.06 % · MortgageTax · Accuracy · Cost In / Out: $5 / $25; Latency: 7.98s [1]	Published observations span 3.08 percentage points. Configuration and source differences contribute to this spread; it is not measured run-to-run uncertainty. #12 of 166 in this displayed comparison; 4.08 pp below the leader.
Qwen 3 Max →Qwen	73.51% 1 report	MortgageTax & TaxEval v2[2]Cost In / Out: $1.2 / $6; Latency: 29.30s	#13 of 166 in this displayed comparison; 4.17 pp below the leader.
Claude Fable 5.1 →Anthropic	73.38% 2 reports · range 70.79–75.96%	MortgageTax & TaxEval v2[2][1] View 2 reports 75.96 % · TaxEval · Accuracy · Cost In / Out: $10 / $50; Latency: 85.56s [2] 70.79 % · MortgageTax · Accuracy · Cost In / Out: $10 / $50; Latency: 14.98s [1]	Published observations span 5.17 percentage points. Configuration and source differences contribute to this spread; it is not measured run-to-run uncertainty. #14 of 166 in this displayed comparison; 4.31 pp below the leader.
GLM 5.2 →Z.ai	73.34% 1 report	MortgageTax & TaxEval v2[2]Cost In / Out: $1.4 / $4.4; Latency: 64.94s	#15 of 166 in this displayed comparison; 4.34 pp below the leader.
Nemotron 3 Ultra →NVIDIA	73.1% 1 report	MortgageTax & TaxEval v2[2]Cost In / Out: N/A; Latency: 36.48s	#16 of 166 in this displayed comparison; 4.58 pp below the leader.
DeepSeek V4 Pro 0813 →DeepSeek	73.06% 1 report	MortgageTax & TaxEval v2[2]Cost In / Out: $1.32 / $3.96; Latency: 3m46s	#17 of 166 in this displayed comparison; 4.62 pp below the leader.
Grok 3 Mini Reasoning →xAICost In / Out: $0.3 / $0.5; Latency: 11.73s	72.98% 1 report	MortgageTax & TaxEval v2[2]Cost In / Out: $0.3 / $0.5; Latency: 11.73s	#18 of 166 in this displayed comparison; 4.7 pp below the leader.
Claude Fable 5 →Anthropic	72.93% 2 reports · range 68.92–76.94%	MortgageTax & TaxEval v2[2][1] View 2 reports 76.94 % · TaxEval · Accuracy · Cost In / Out: $10 / $50; Latency: 56.87s [2] 68.92 % · MortgageTax · Accuracy · Cost In / Out: $10 / $50; Latency: 16.10s [1]	Published observations span 8.02 percentage points. Configuration and source differences contribute to this spread; it is not measured run-to-run uncertainty. #19 of 166 in this displayed comparison; 4.75 pp below the leader.
Muse Spark 1.1 →Meta	72.93% 2 reports · range 66.14–79.72%	MortgageTax & TaxEval v2[2][1] View 2 reports 79.72 % · TaxEval · Accuracy · Cost In / Out: $1.25 / $4.25; Latency: 32.82s [2] 66.14 % · MortgageTax · Accuracy · Cost In / Out: $1.25 / $4.25; Latency: 24.82s [1]	Published observations span 13.58 percentage points. Configuration and source differences contribute to this spread; it is not measured run-to-run uncertainty. #19 of 166 in this displayed comparison; 4.75 pp below the leader.
Muse Spark 1.2 →Meta	72.9% 2 reports · range 65.42–80.38%	MortgageTax & TaxEval v2[2][1] View 2 reports 80.38 % · TaxEval · Accuracy · Cost In / Out: $1.25 / $4.25; Latency: 48.31s [2] 65.42 % · MortgageTax · Accuracy · Cost In / Out: $1.25 / $4.25; Latency: 44.62s [1]	Published observations span 14.96 percentage points. Configuration and source differences contribute to this spread; it is not measured run-to-run uncertainty. #21 of 166 in this displayed comparison; 4.78 pp below the leader.
Claude Sonnet 5 →Anthropic	72.83% 2 reports · range 70.03–75.63%	MortgageTax & TaxEval v2[2][1] View 2 reports 75.63 % · TaxEval · Accuracy · Cost In / Out: $2 / $10; Latency: 3m22s [2] 70.03 % · MortgageTax · Accuracy · Cost In / Out: $2 / $10; Latency: 28.67s [1]	Published observations span 5.6 percentage points. Configuration and source differences contribute to this spread; it is not measured run-to-run uncertainty. #22 of 166 in this displayed comparison; 4.85 pp below the leader.
Claude Opus 4.7 →Anthropic	72.77% 2 reports · range 70.27–75.27%	MortgageTax & TaxEval v2[2][1] View 2 reports 75.27 % · TaxEval · Accuracy · Cost In / Out: $5 / $25; Latency: 62.07s [2] 70.27 % · MortgageTax · Accuracy · Cost In / Out: $5 / $25; Latency: 16.10s [1]	Published observations span 5 percentage points. Configuration and source differences contribute to this spread; it is not measured run-to-run uncertainty. #23 of 166 in this displayed comparison; 4.91 pp below the leader.
Claude Opus 4.8 →Anthropic	72.77% 2 reports · range 69.91–75.63%	MortgageTax & TaxEval v2[2][1] View 2 reports 75.63 % · TaxEval · Accuracy · Cost In / Out: $5 / $25; Latency: 70.33s [2] 69.91 % · MortgageTax · Accuracy · Cost In / Out: $5 / $25; Latency: 20.81s [1]	Published observations span 5.72 percentage points. Configuration and source differences contribute to this spread; it is not measured run-to-run uncertainty. #23 of 166 in this displayed comparison; 4.91 pp below the leader.
Gemini 2.5 Flash Preview (9/25) →GoogleCost In / Out: $0.3 / $2.5; Latency: 6.27s	72.73% 1 report	MortgageTax & TaxEval v2[2]Cost In / Out: $0.3 / $2.5; Latency: 6.27s	#25 of 166 in this displayed comparison; 4.95 pp below the leader.
Claude Sonnet 4.6 →Anthropic	72.42% 2 reports · range 67.73–77.11%	MortgageTax & TaxEval v2[2][1] View 2 reports 77.11 % · TaxEval · Accuracy · Cost In / Out: $3 / $15; Latency: 2m08s [2] 67.73 % · MortgageTax · Accuracy · Cost In / Out: $3 / $15; Latency: 29.74s [1]	Published observations span 9.38 percentage points. Configuration and source differences contribute to this spread; it is not measured run-to-run uncertainty. #26 of 166 in this displayed comparison; 5.26 pp below the leader.
Claude 3.7 Sonnet →AnthropicCost In / Out: $3 / $15; Latency: 6.28s	72.4% 1 report	MortgageTax & TaxEval v2[2]Cost In / Out: $3 / $15; Latency: 6.28s	#27 of 166 in this displayed comparison; 5.28 pp below the leader.
Gemini 2.5 Flash Preview (9/25) →GoogleCost In / Out: $0.3 / $2.5; Latency: 10.63s	72.4% 1 report	MortgageTax & TaxEval v2[2]Cost In / Out: $0.3 / $2.5; Latency: 10.63s	#27 of 166 in this displayed comparison; 5.28 pp below the leader.
GLM 4.5 →Z.ai	72.4% 1 report	MortgageTax & TaxEval v2[2]Cost In / Out: $0.6 / $2.2; Latency: 2m09s	#27 of 166 in this displayed comparison; 5.28 pp below the leader.
GLM 5.3 →Z.ai	72.36% 1 report	MortgageTax & TaxEval v2[2]Cost In / Out: $1.4 / $4.4; Latency: 98.32s	#30 of 166 in this displayed comparison; 5.32 pp below the leader.
DeepSeek R1 →DeepSeek	72.28% 1 report	MortgageTax & TaxEval v2[2]Cost In / Out: $1.35 / $5.4; Latency: 2m33s	#31 of 166 in this displayed comparison; 5.4 pp below the leader.
Claude Opus 4.6 →Anthropic	72.24% 2 reports · range 68.52–75.96%	MortgageTax & TaxEval v2[2][1] View 2 reports 75.96 % · TaxEval · Accuracy · Cost In / Out: $5 / $25; Latency: 79.25s [2] 68.52 % · MortgageTax · Accuracy · Cost In / Out: $5 / $25; Latency: 27.16s [1]	Published observations span 7.44 percentage points. Configuration and source differences contribute to this spread; it is not measured run-to-run uncertainty. #32 of 166 in this displayed comparison; 5.44 pp below the leader.
DeepSeek V4 →DeepSeek	72.08% 1 report	MortgageTax & TaxEval v2[2]Cost In / Out: $1.32 / $3.96; Latency: 3m10s	#33 of 166 in this displayed comparison; 5.6 pp below the leader.
Claude Sonnet 4 →AnthropicCost In / Out: $3 / $15; Latency: 32.38s	72% 1 report	MortgageTax & TaxEval v2[2]Cost In / Out: $3 / $15; Latency: 32.38s	#34 of 166 in this displayed comparison; 5.68 pp below the leader.
GPT-5.5 →OpenAI	71.87% 2 reports · range 68.76–74.98%	MortgageTax & TaxEval v2[2][1] View 2 reports 74.98 % · TaxEval · Accuracy · Cost In / Out: $5 / $30; Latency: 95.83s [2] 68.76 % · MortgageTax · Accuracy · Cost In / Out: $5 / $30; Latency: 28.24s [1]	Published observations span 6.22 percentage points. Configuration and source differences contribute to this spread; it is not measured run-to-run uncertainty. #35 of 166 in this displayed comparison; 5.81 pp below the leader.
GPT-5.6 Terra →OpenAI	71.75% 2 reports · range 67.33–76.17%	MortgageTax & TaxEval v2[2][1] View 2 reports 76.17 % · TaxEval · Accuracy · Cost In / Out: $2 / $12; Latency: 37.28s [2] 67.33 % · MortgageTax · Accuracy · Cost In / Out: $2 / $12; Latency: 6.32s [1]	Published observations span 8.84 percentage points. Configuration and source differences contribute to this spread; it is not measured run-to-run uncertainty. #36 of 166 in this displayed comparison; 5.93 pp below the leader.
GPT-5.6 Luna →OpenAI	71.73% 2 reports · range 67.29–76.17%	MortgageTax & TaxEval v2[2][1] View 2 reports 76.17 % · TaxEval · Accuracy · Cost In / Out: $0.2 / $1.2; Latency: 76.52s [2] 67.29 % · MortgageTax · Accuracy · Cost In / Out: $0.2 / $1.2; Latency: 22.84s [1]	Published observations span 8.88 percentage points. Configuration and source differences contribute to this spread; it is not measured run-to-run uncertainty. #37 of 166 in this displayed comparison; 5.95 pp below the leader.
Kimi K2 Thinking →Kimi	71.71% 1 report	MortgageTax & TaxEval v2[2]Cost In / Out: $0.6 / $2.5; Latency: 61.41s	#38 of 166 in this displayed comparison; 5.97 pp below the leader.
GPT OSS 120B →OpenAI	71.59% 1 report	MortgageTax & TaxEval v2[2]Cost In / Out: $0.15 / $0.6; Latency: 47.66s	#39 of 166 in this displayed comparison; 6.09 pp below the leader.
Claude Opus 4.1 →AnthropicCost In / Out: $15 / $75; Latency: 32.60s	71.46% 1 report	MortgageTax & TaxEval v2[2]Cost In / Out: $15 / $75; Latency: 32.60s	#40 of 166 in this displayed comparison; 6.22 pp below the leader.
GPT-5.2 →OpenAI	71.44% 2 reports · range 67.13–75.76%	MortgageTax & TaxEval v2[2][1] View 2 reports 75.76 % · TaxEval · Accuracy · Cost In / Out: $1.75 / $14; Latency: 60.48s [2] 67.13 % · MortgageTax · Accuracy · Cost In / Out: $1.75 / $14; Latency: 70.67s [1]	Published observations span 8.63 percentage points. Configuration and source differences contribute to this spread; it is not measured run-to-run uncertainty. #41 of 166 in this displayed comparison; 6.24 pp below the leader.
Qwen 3.6 Plus →Qwen	71.35% 2 reports · range 67.97–74.73%	MortgageTax & TaxEval v2[2][1] View 2 reports 74.73 % · TaxEval · Accuracy · Cost In / Out: $0.5 / $3; Latency: 70.63s [2] 67.97 % · MortgageTax · Accuracy · Cost In / Out: $0.5 / $3; Latency: 78.18s [1]	Published observations span 6.76 percentage points. Configuration and source differences contribute to this spread; it is not measured run-to-run uncertainty. #42 of 166 in this displayed comparison; 6.33 pp below the leader.
Gemini 3 Flash (12/25) →Google	71.3% 2 reports · range 68.72–73.88%	MortgageTax & TaxEval v2[2][1] View 2 reports 73.88 % · TaxEval · Accuracy · Cost In / Out: $0.5 / $3; Latency: 12.00s [2] 68.72 % · MortgageTax · Accuracy · Cost In / Out: $0.5 / $3; Latency: 12.84s [1]	Published observations span 5.16 percentage points. Configuration and source differences contribute to this spread; it is not measured run-to-run uncertainty. #43 of 166 in this displayed comparison; 6.38 pp below the leader.
Gemini 3.5 Flash →Google	71.25% 2 reports · range 68.12–74.37%	MortgageTax & TaxEval v2[2][1] View 2 reports 74.37 % · TaxEval · Accuracy · Cost In / Out: $1.5 / $9; Latency: 19.69s [2] 68.12 % · MortgageTax · Accuracy · Cost In / Out: $1.5 / $9; Latency: 11.43s [1]	Published observations span 6.25 percentage points. Configuration and source differences contribute to this spread; it is not measured run-to-run uncertainty. #44 of 166 in this displayed comparison; 6.44 pp below the leader.
Gemini 3.6 Flash →Google	71.23% 2 reports · range 67.61–74.86%	MortgageTax & TaxEval v2[2][1] View 2 reports 74.86 % · TaxEval · Accuracy · Cost In / Out: $1.5 / $7.5; Latency: 12.34s [2] 67.61 % · MortgageTax · Accuracy · Cost In / Out: $1.5 / $7.5; Latency: 9.24s [1]	Published observations span 7.25 percentage points. Configuration and source differences contribute to this spread; it is not measured run-to-run uncertainty. #45 of 166 in this displayed comparison; 6.45 pp below the leader.
GLM 5.1 →Z.ai	71.19% 1 report	MortgageTax & TaxEval v2[2]Cost In / Out: $1 / $3.2; Latency: 84.07s	#46 of 166 in this displayed comparison; 6.49 pp below the leader.
Gemini 2.5 Flash Preview 4/17 →GoogleCost In / Out: $0.3 / $2.5; Latency: 5.39s	71.18% 1 report	MortgageTax & TaxEval v2[2]Cost In / Out: $0.3 / $2.5; Latency: 5.39s	#47 of 166 in this displayed comparison; 6.5 pp below the leader.
Gemini 3.1 Pro Preview (02/26) →Google	71.14% 2 reports · range 69.4–72.88%	MortgageTax & TaxEval v2[2][1] View 2 reports 72.88 % · TaxEval · Accuracy · Cost In / Out: $2 / $12; Latency: 41.93s [2] 69.4 % · MortgageTax · Accuracy · Cost In / Out: $2 / $12; Latency: 23.62s [1]	Published observations span 3.48 percentage points. Configuration and source differences contribute to this spread; it is not measured run-to-run uncertainty. #48 of 166 in this displayed comparison; 6.54 pp below the leader.
Grok 3 Mini Reasoning →xAICost In / Out: $0.3 / $0.5; Latency: 25.45s	71.14% 1 report	MortgageTax & TaxEval v2[2]Cost In / Out: $0.3 / $0.5; Latency: 25.45s	#48 of 166 in this displayed comparison; 6.54 pp below the leader.
GPT-5.4 →OpenAI	71.14% 2 reports · range 68.32–73.96%	MortgageTax & TaxEval v2[2][1] View 2 reports 73.96 % · TaxEval · Accuracy · Cost In / Out: $2.5 / $15; Latency: 78.72s [2] 68.32 % · MortgageTax · Accuracy · Cost In / Out: $2.5 / $15; Latency: 45.00s [1]	Published observations span 5.64 percentage points. Configuration and source differences contribute to this spread; it is not measured run-to-run uncertainty. #50 of 166 in this displayed comparison; 6.54 pp below the leader.

Sources

MortgageTax — evaluation methodology and results ↗
Vals AI · 2026
Dates & evaluation details
MortgageTax · Accuracy · 2026-10-26
Public source leaderboard captured 26 October 2026. Model names, configurations, native score units and source-reported uncertainty are retained. Version and task split follow this specific Vals methodology page. These are published observations, not evaluations run by RLTasks.
TaxEval — evaluation methodology and results ↗
Vals AI · 2026
Dates & evaluation details
TaxEval · Accuracy · 2026-10-26
Public source leaderboard captured 26 October 2026. Model names, configurations, native score units and source-reported uncertainty are retained. Version and task split follow this specific Vals methodology page. These are published observations, not evaluations run by RLTasks.

TRACE THE SCORE · Per-model notes use published breakdowns where available.

What it measures

Visual tax-certificate extraction (MortgageTax) and tax regulation QA (TaxEval v2).

Reading the measurement

Check the source documents, dates, accounting conventions and calculation chain. A financial answer needs evidence and consistent units. Benchmark success does not establish live investment performance or suitability for an individual decision.

Published evaluation

The linked publication documents the task and scoring procedure. Model results retain the source’s evaluation setting, native score units and available uncertainty.

Where this fits in an evaluation suite

Synthetic research rewards do not establish live investment performance. Preserve transaction costs, held-out regimes, constraints, and the exact policy behind a reported result.

Start with a capability hypothesis, then choose a second test with a different failure mode. Keep the checkpoint, reasoning effort, agent tools, sampling policy, and scoring procedure fixed when measuring a training intervention. Record failures and missing outputs in the denominator.

A reproducible comparison

Record the source publication, benchmark version, scoring units, denominator, tool access and attempt budget. Preserve uncertainty and distinguish developer-reported results from independent evaluations.

Publish the exact run configuration, item coverage, score units, and available uncertainty. If a source supplies only an aggregate or a single run, preserve that limit. A new model name or a rounded score alone does not establish a capability improvement.

PRIMARY EVIDENCE & FURTHER READING

MortgageTax — evaluation methodology and results ↗
Vals AI · 2026
TaxEval — evaluation methodology and results ↗
Vals AI · 2026

Reviewed 2026-10-26. Explanation and interpretation by RLTasks. Benchmark ownership remains with its original authors.

Continue exploring

All financial reasoning benchmarks →

Sources

What it measures

Reading the measurement

Published evaluation

Where this fits in an evaluation suite

A reproducible comparison

Continue exploring

τ³-Banking

BankerToolBench

OfficeQA Pro

Vals Finance Agent v2