# Vals AI – Compare AI models across real-world benchmarks

> Source: <https://www.vals.ai/comparison>
> Published: 2026-11-19 05:28:34+00:00

**Recommendation:** Evaluate models per task domain rather than by aggregate index: Gemini 4 Neon beat Claude Fable 5.2 by 10.1 points on Legal Research Bench and 22.8 points on Tax Agent Bench despite trailing on the overall Vals Index.
More in the playbook "Evals and testing for LLM apps": <https://wpnews.pro/playbooks/evals-testing>

# Compare Models

Select models (max 7)

Claude Fable 5.2Gemini 4 NeonClaude Sonnet 5.5Gemini 4 CarbonClaude Opus 5.5

| Vals IndexGDP-weighted benchmark | 71.31%±0.96 | 70.87%±0.94 | 70.85%±0.93 | 70.45%±0.97 | 70.14%±1.00 | 
| Legal Research BenchAgentic US legal research | 46.15%±3.46 | 56.25%±3.45 | 48.08%±3.47 | 55.77%±3.45 | 50.48%±3.48 | 
| Finance Agent v2.1Core financial analyst tasks | 69.47%±1.98 | 68.56%±1.94 | 67.28%±2.02 | 67.16%±2.02 | 67.85%±2.05 | 
| Tax Agent BenchAgentic US corporate tax research | 32.09%±3.37 | 54.89%±3.59 | 41.96%±3.56 | 51.78%±3.60 | 44.54%±3.58 | 
| MedCodeMedical billing code support | N/A | N/A | 52.92%±2.12 | N/A | 49.80%±2.27 | 
| Terminal-Bench-ScienceExpert-authored scientific research workflows | 55.71%±5.98 | 62.86%±5.82 | 47.14%±6.01 | N/A | 47.14%±6.01 | 
| Code Migration v1.1Rewriting programs in new languages | 83.27%±3.12 | N/A | 86.88%±1.52 | N/A | 81.01%±3.21 | 
| Terminal-Bench 4.0Frontier-difficulty terminal tasks | 65.15%±2.31 | 61.62%±2.02 | 64.14%±1.01 | 63.13%±2.20 | 65.15%±0.00 | 
| Vibe Code Bench v1.1Building web apps from scratch | 93.77%±1.15 | 93.70%±1.17 | 92.39%±1.26 | 89.06%±2.29 | 90.29%±1.53 | 

Benchmarks

Vals Index *

Claude Fable 5.2

0.00%± 0.96

(42/42)
Gemini 4 Neon

0.00%± 0.94

(42/42)
Claude Sonnet 5.5

0.00%± 0.93

(42/42)
Gemini 4 Carbon

0.00%± 0.97

(42/42)
Claude Opus 5.5

0.00%± 1.00

(42/42)
Legal Research Bench *

Claude Fable 5.2

0.00%± 3.46

(99/99)
Gemini 4 Neon

0.00%± 3.45

(99/99)
Claude Sonnet 5.5

0.00%± 3.47

(99/99)
Gemini 4 Carbon

0.00%± 3.45

(99/99)
Claude Opus 5.5

0.00%± 3.48

(99/99)
Finance Agent v2.1 *

Claude Fable 5.2

0.00%± 1.98

(45/45)
Gemini 4 Neon

0.00%± 1.94

(45/45)
Claude Sonnet 5.5

0.00%± 2.02

(45/45)
Gemini 4 Carbon

0.00%± 2.02

(45/45)
Claude Opus 5.5

0.00%± 2.05

(45/45)
Tax Agent Bench *

Claude Fable 5.2

0.00%± 3.37

(90/90)
Gemini 4 Neon

0.00%± 3.59

(90/90)
Claude Sonnet 5.5

0.00%± 3.56

(90/90)
Gemini 4 Carbon

0.00%± 3.60

(90/90)
Claude Opus 5.5

0.00%± 3.58

(90/90)
MedCode *

Claude Fable 5.2

N/A
Gemini 4 Neon

N/A
Claude Sonnet 5.5

0.00%± 2.12

(108/108)
Gemini 4 Carbon

N/A
Claude Opus 5.5

0.00%± 2.27

(108/108)
Terminal-Bench-Science

Claude Fable 5.2

0.00%± 5.98

(48/48)
Gemini 4 Neon

0.00%± 5.82

(48/48)
Claude Sonnet 5.5

0.00%± 6.01

(48/48)
Gemini 4 Carbon

N/A
Claude Opus 5.5

0.00%± 6.01

(48/48)
Code Migration v1.1 *

Claude Fable 5.2

0.00%± 3.12

(30/30)
Gemini 4 Neon

N/A
Claude Sonnet 5.5

0.00%± 1.52

(30/30)
Gemini 4 Carbon

N/A
Claude Opus 5.5

0.00%± 3.21

(30/30)
Terminal-Bench 4.0

Claude Fable 5.2

0.00%± 2.31

(69/69)
Gemini 4 Neon

0.00%± 2.02

(69/69)
Claude Sonnet 5.5

0.00%± 1.01

(69/69)
Gemini 4 Carbon

0.00%± 2.20

(69/69)
Claude Opus 5.5

0.00%± 0.00

(69/69)
Vibe Code Bench v1.1 *

Claude Fable 5.2

0.00%± 1.15

(133/133)
Gemini 4 Neon

0.00%± 1.17

(133/133)
Claude Sonnet 5.5

0.00%± 1.26

(133/133)
Gemini 4 Carbon

0.00%± 2.29

(133/133)
Claude Opus 5.5

0.00%± 1.53

(133/133)
## Overall performance

Performance on the Vals Index, a GDP-weighted aggregation of tasks across finance, coding, legal, and tax

•2026-11-18

## Comparison by Industry

Model performance on different sections of the economy.

Legal

41.71%

vs.
37.92%

vs.
42.49%

vs.
36.63%

vs.
43.93%

Finance

59.65%

vs.
66.64%

vs.
71.28%

vs.
65.14%

vs.
72.05%

Healthcare

N/A

vs.
N/A

vs.
72.01%

vs.
N/A

vs.
70.61%

Math

92.86%

vs.
N/A

vs.
100.00%

vs.
71.43%

vs.
100.00%

Science

63.74%

vs.
62.63%

vs.
59.12%

vs.
N/A

vs.
47.63%

Education

N/A

vs.
N/A

vs.
88.03%

vs.
N/A

vs.
91.30%

Coding

64.70%

vs.
77.66%

vs.
65.22%

vs.
67.31%

vs.
64.45%

Cyber

N/A

vs.
N/A

vs.
44.87%

vs.
N/A

vs.
55.35%

Social Mobility

N/A

vs.
N/A

vs.
72.36%

vs.
N/A

vs.
73.52%

## Cost Analysis

Per task and per token model pricing.

Vals Index · USD

$17.94

$7.405

$16.18

$10.51

$24.77

Cost / TestVals Index

$17.94

vs.
$7.405

vs.
$16.18

vs.
$10.51

vs.
$24.77

Input Cost/ 1M Tokens

$8.00

vs.
$3.00

vs.
$2.00

vs.
$4.00

vs.
$4.00

Input Cache Write/ 1M Tokens

$10.00

vs.
N/A

vs.
$2.50

vs.
N/A

vs.
$5.00

Input Cache Read/ 1M Tokens

$0.20

vs.
$0.15

vs.
$0.10

vs.
$0.20

vs.
$0.20

Output Cost/ 1M Tokens

$40.00

vs.
$15.00

vs.
$10.00

vs.
$20.00

vs.
$20.00

## Model Metadata

Basic information about each model.

| Model provider | Anthropic | Google | Anthropic | Google | Anthropic | 
| Latency | 2621.82s | 1884.68s | 5158.77s | 2864.86s | 5071.09s | 
| Cost (In/Out) | $8 / $40 | $3 / $15 | $2 / $10 | $4 / $20 | $4 / $20 | 
| Context Window | 1M | 1M | 1M | 1M | 1M | 
| Max Output Token | 128k | 1M | 128k | 1M | 128k | 
| Input Modality |  |  |  |  |  |
