Dataset Viewer
Auto-converted to Parquet Duplicate
task
stringclasses
8 values
difficulty
stringclasses
3 values
rank
int64
1
4
model
stringclasses
13 values
lab
stringclasses
7 values
quality_0_100
int64
69
98
expected_cost_usd
float64
0
0.66
expected_seconds
float64
1.4
124
thinking_effort
stringclasses
4 values
updated
stringdate
2026-11-20 00:00:00
2026-11-20 00:00:00
quick answers
Easy
1
GPT-6 Luna
OpenAI
95
0.000105
1.5
none
2026-11-20
quick answers
Easy
2
Claude Haiku
Anthropic
88
0.000105
1.8
none
2026-11-20
quick answers
Easy
3
DeepSeek Flash
DeepSeek
92
0.00027
2.2
none
2026-11-20
quick answers
Easy
4
Gemini Flash
Google
92
0.000788
1.4
none
2026-11-20
quick answers
Medium
1
GPT-6 Luna
OpenAI
95
0.000105
1.5
none
2026-11-20
quick answers
Medium
2
Claude Haiku
Anthropic
88
0.000105
1.8
none
2026-11-20
quick answers
Medium
3
Gemini Flash
Google
92
0.000788
1.4
none
2026-11-20
quick answers
Medium
4
DeepSeek Flash
DeepSeek
92
0.00027
2.2
none
2026-11-20
quick answers
Hard
1
GPT-6 Luna
OpenAI
95
0.000255
3.5
low
2026-11-20
quick answers
Hard
2
Gemini Flash
Google
92
0.001912
3.3
low
2026-11-20
quick answers
Hard
3
DeepSeek Flash
DeepSeek
92
0.00063
5.2
low
2026-11-20
quick answers
Hard
4
GPT-6.1 Sol
OpenAI
93
0.0051
7.4
low
2026-11-20
writing
Easy
1
GPT-6 Luna
OpenAI
89
0.00037
3.8
none
2026-11-20
writing
Easy
2
Claude Haiku
Anthropic
88
0.00037
4.7
none
2026-11-20
writing
Easy
3
DeepSeek Flash
DeepSeek
85
0.00096
5.7
none
2026-11-20
writing
Easy
4
Gemini Flash
Google
80
0.002775
3.6
none
2026-11-20
writing
Medium
1
GPT-6 Luna
OpenAI
89
0.00037
3.8
none
2026-11-20
writing
Medium
2
Claude Haiku
Anthropic
88
0.00037
4.7
none
2026-11-20
writing
Medium
3
Claude Sonnet
Anthropic
93
0.0074
8
none
2026-11-20
writing
Medium
4
GPT-6.1 Sol
OpenAI
92
0.0074
8.1
none
2026-11-20
writing
Hard
1
Claude Opus
Anthropic
96
0.0208
19.2
low
2026-11-20
writing
Hard
2
Claude Sonnet
Anthropic
93
0.0104
12.3
low
2026-11-20
writing
Hard
3
GPT-5.5
OpenAI
93
0.03
17.5
low
2026-11-20
writing
Hard
4
GPT-6.1 Sol
OpenAI
92
0.0104
12.4
low
2026-11-20
coding
Easy
1
GPT-6 Luna
OpenAI
87
0.0005
3.8
none
2026-11-20
coding
Easy
2
Claude Haiku
Anthropic
85
0.0005
4.7
none
2026-11-20
coding
Easy
3
DeepSeek Flash
DeepSeek
88
0.00135
5.7
none
2026-11-20
coding
Easy
4
Gemini Flash
Google
81
0.00375
3.6
none
2026-11-20
coding
Medium
1
Claude Sonnet
Anthropic
94
0.013
12.3
low
2026-11-20
coding
Medium
2
GPT-6.1 Sol
OpenAI
94
0.013
12.4
low
2026-11-20
coding
Medium
3
Kimi K3.1 Pro
Moonshot
92
0.013
14.4
low
2026-11-20
coding
Medium
4
Claude Opus
Anthropic
98
0.026
19.2
low
2026-11-20
coding
Hard
1
Claude Opus
Anthropic
98
0.12
123.6
high
2026-11-20
coding
Hard
2
Claude Sonnet
Anthropic
94
0.06
79.5
high
2026-11-20
coding
Hard
3
GPT-6.1 Sol
OpenAI
94
0.06
79.6
high
2026-11-20
coding
Hard
4
GPT-5.5
OpenAI
92
0.1775
111.5
high
2026-11-20
reasoning
Easy
1
GPT-6 Luna
OpenAI
78
0.0004
3.8
none
2026-11-20
reasoning
Easy
2
Claude Haiku
Anthropic
73
0.0004
4.7
none
2026-11-20
reasoning
Easy
3
DeepSeek Flash
DeepSeek
78
0.00105
5.7
none
2026-11-20
reasoning
Easy
4
Gemini Flash
Google
73
0.003
3.6
none
2026-11-20
reasoning
Medium
1
Claude Sonnet
Anthropic
90
0.023
29.5
medium
2026-11-20
reasoning
Medium
2
Claude Opus
Anthropic
98
0.046
45.8
medium
2026-11-20
reasoning
Medium
3
Kimi K3.1 Pro
Moonshot
86
0.023
34.4
low
2026-11-20
reasoning
Medium
4
Muse Spark
Meta
83
0.010375
18.8
medium
2026-11-20
reasoning
Hard
1
Claude Opus
Anthropic
98
0.116
123.6
high
2026-11-20
reasoning
Hard
2
Claude Sonnet
Anthropic
90
0.058
79.5
high
2026-11-20
reasoning
Hard
3
Kimi K3.1 Pro
Moonshot
86
0.058
92.8
high
2026-11-20
reasoning
Hard
4
Muse Spark
Meta
83
0.02525
50.6
high
2026-11-20
maths
Easy
1
Claude Haiku
Anthropic
82
0.00048
7.2
low
2026-11-20
maths
Easy
2
GPT-6 Luna
OpenAI
69
0.00048
5.8
low
2026-11-20
maths
Easy
3
Gemini Flash
Google
83
0.0036
5.5
low
2026-11-20
maths
Easy
4
DeepSeek Flash
DeepSeek
76
0.0012
8.7
low
2026-11-20
maths
Medium
1
Mistral Large
Mistral AI
88
0.00476
25.8
none
2026-11-20
maths
Medium
2
Claude Opus
Anthropic
96
0.0432
45.8
medium
2026-11-20
maths
Medium
3
Gemini 4 Argon
Google
86
0.0216
29.8
medium
2026-11-20
maths
Medium
4
DeepSeek V4 Pro
DeepSeek
84
0.008976
41.5
low
2026-11-20
maths
Hard
1
Claude Opus
Anthropic
96
0.1132
123.6
high
2026-11-20
maths
Hard
2
Mistral Large
Mistral AI
88
0.01211
69.5
high
2026-11-20
maths
Hard
3
Gemini 4 Argon
Google
86
0.0566
79.8
high
2026-11-20
maths
Hard
4
DeepSeek V4 Pro
DeepSeek
84
0.022836
111.5
high
2026-11-20
structured data
Easy
1
GPT-6 Luna
OpenAI
92
0.00055
3.8
none
2026-11-20
structured data
Easy
2
Claude Haiku
Anthropic
85
0.00055
4.7
none
2026-11-20
structured data
Easy
3
Gemini Flash
Google
93
0.004125
3.6
none
2026-11-20
structured data
Easy
4
DeepSeek Flash
DeepSeek
91
0.0015
5.7
none
2026-11-20
structured data
Medium
1
GPT-6 Luna
OpenAI
92
0.0007
5.8
low
2026-11-20
structured data
Medium
2
Claude Haiku
Anthropic
85
0.0007
7.2
low
2026-11-20
structured data
Medium
3
Gemini Flash
Google
93
0.00525
5.5
low
2026-11-20
structured data
Medium
4
DeepSeek Flash
DeepSeek
91
0.00186
8.7
low
2026-11-20
structured data
Hard
1
GPT-6 Luna
OpenAI
92
0.0013
13.8
medium
2026-11-20
structured data
Hard
2
Gemini Flash
Google
93
0.00975
13
medium
2026-11-20
structured data
Hard
3
DeepSeek Flash
DeepSeek
91
0.0033
20.7
low
2026-11-20
structured data
Hard
4
Gemini 4 Argon
Google
92
0.026
29.8
medium
2026-11-20
long documents
Easy
1
GPT-6 Luna
OpenAI
91
0.01225
3.8
none
2026-11-20
long documents
Easy
2
Gemini Flash
Google
96
0.091875
3.6
none
2026-11-20
long documents
Easy
3
DeepSeek Flash
DeepSeek
93
0.0366
5.7
none
2026-11-20
long documents
Easy
4
Claude Haiku
Anthropic
90
0.06125
4.7
none
2026-11-20
long documents
Medium
1
GPT-6 Luna
OpenAI
91
0.01225
3.8
none
2026-11-20
long documents
Medium
2
Gemini Flash
Google
96
0.091875
3.6
none
2026-11-20
long documents
Medium
3
DeepSeek Flash
DeepSeek
93
0.0366
5.7
none
2026-11-20
long documents
Medium
4
Claude Haiku
Anthropic
90
0.06125
4.7
none
2026-11-20
long documents
Hard
1
Gemini Flash
Google
96
0.0975
13
medium
2026-11-20
long documents
Hard
2
DeepSeek Flash
DeepSeek
93
0.0384
20.7
low
2026-11-20
long documents
Hard
3
GPT-6.1 Sol
OpenAI
93
0.26
29.6
medium
2026-11-20
long documents
Hard
4
GPT-5.5
OpenAI
94
0.66
41.5
medium
2026-11-20
translation
Easy
1
GPT-6 Luna
OpenAI
92
0.00035
3.8
none
2026-11-20
translation
Easy
2
Claude Haiku
Anthropic
86
0.00035
4.7
none
2026-11-20
translation
Easy
3
Gemini Flash
Google
93
0.002625
3.6
none
2026-11-20
translation
Easy
4
DeepSeek Flash
DeepSeek
92
0.0009
5.7
none
2026-11-20
translation
Medium
1
GPT-6 Luna
OpenAI
92
0.00035
3.8
none
2026-11-20
translation
Medium
2
Claude Haiku
Anthropic
86
0.00035
4.7
none
2026-11-20
translation
Medium
3
Gemini Flash
Google
93
0.002625
3.6
none
2026-11-20
translation
Medium
4
DeepSeek Flash
DeepSeek
92
0.0009
5.7
none
2026-11-20
translation
Hard
1
GPT-6 Luna
OpenAI
92
0.0005
5.8
low
2026-11-20
translation
Hard
2
Gemini Flash
Google
93
0.00375
5.5
low
2026-11-20
translation
Hard
3
DeepSeek Flash
DeepSeek
92
0.00126
8.7
low
2026-11-20
translation
Hard
4
Muse Spark
Meta
91
0.00465
7.9
low
2026-11-20

The Router Index

Which AI model is best for each kind of task, measured. The Router Index is what Nivana's model router picks for 8 kinds of request at 3 difficulties. The picks come from measured quality, live prices and speed, and the data is refreshed monthly.

What the router picks

Task Difficulty Picked Runner-up
quick answers easy GPT-6 Luna Claude Haiku
quick answers medium GPT-6 Luna Claude Haiku
quick answers hard GPT-6 Luna Gemini Flash
writing easy GPT-6 Luna Claude Haiku
writing medium GPT-6 Luna Claude Haiku
writing hard Claude Opus Claude Sonnet
coding easy GPT-6 Luna Claude Haiku
coding medium Claude Sonnet GPT-6.1 Sol
coding hard Claude Opus Claude Sonnet
reasoning easy GPT-6 Luna Claude Haiku
reasoning medium Claude Sonnet Claude Opus
reasoning hard Claude Opus Claude Sonnet
maths easy Claude Haiku GPT-6 Luna
maths medium Mistral Large Claude Opus
maths hard Claude Opus Mistral Large
structured data easy GPT-6 Luna Claude Haiku
structured data medium GPT-6 Luna Claude Haiku
structured data hard GPT-6 Luna Gemini Flash
long documents easy GPT-6 Luna Gemini Flash
long documents medium GPT-6 Luna Gemini Flash
long documents hard Gemini Flash DeepSeek Flash
translation easy GPT-6 Luna Claude Haiku
translation medium GPT-6 Luna Claude Haiku
translation hard GPT-6 Luna Gemini Flash

Measured quality by task (0–100)

Model Lab Quick answers Writing Coding Reasoning Maths Structured data Long documents Translation Price in / out per 1M tokens
Claude Opus Anthropic 83 96 98 98 96 88 88 93 $4 / $20
Claude Sonnet Anthropic 85 93 94 90 84 88 90 86 $2 / $10
Claude Haiku Anthropic 88 88 85 73 82 85 90 86 $0.1 / $0.5
GPT-6.1 Sol OpenAI 93 92 94 82 76 90 93 91 $2 / $10
GPT-6 Luna OpenAI 95 89 87 78 69 92 91 92 $0.1 / $0.5
GPT-5.5 OpenAI 94 93 92 78 78 94 94 94 $5 / $30
Gemini 4 Argon Google 91 81 81 80 86 92 89 90 $2 / $10
Gemini Flash Google 92 80 81 73 83 93 96 93 $0.75 / $3.75
DeepSeek V4 Pro DeepSeek 91 88 83 75 84 92 90 91 $1.32 / $3.96
DeepSeek Flash DeepSeek 92 85 88 78 76 91 93 92 $0.3 / $1.2
Mistral Large Mistral AI 84 84 78 75 88 88 91 91 $0.7 / $2.1
Muse Spark Meta 91 85 82 83 81 87 92 91 $1.25 / $4.25
Kimi K3.1 Pro Moonshot 92 92 92 86 79 92 92 92 $2 / $10
Grok 4.7 xAI 91 88 88 81 80 87 89 89 $2 / $6

Does routing beat one strong model?

The routing benchmark compares Nivana's automatic routing with always calling Claude Opus 5.5 or Sonnet 5.5, on 96 held-out prompts in two sets. Three judges from other labs compared each pair in both orders. On 2026-11-03, the judges preferred Opus over Nivana: Nivana scored 44% (95% CI 38–49%), and a caller paid 111% of the Opus price. Against Sonnet, Nivana was level (51%) at 2.6 times the price. Routing paid off on easy requests, which were level with Opus at a quarter of the cost. It did worst on hard requests and on writing. Prompts, answers, verdicts and the harness are all in benchmark/, so anyone can run it again.

Method

Quality from weekly pairwise evals per task judged by models from several labs (0–100); cost from live AI Gateway list prices for a typical request of that kind; speed from live traffic.

The best pick depends on the request, not just the top score. A few points of quality rarely show on an easy request, while a 10–100Γ— price difference always does. The router therefore weighs cost and speed heavily for easy requests and almost ignores them for hard ones. Full write-up: https://nivana.one/router-index#method.

Columns (CSV)

task, difficulty, rank, model, lab, quality_0_100, expected_cost_usd, expected_seconds, thinking_effort, updated. Each row is one candidate in the router's ranking for a task and difficulty. Rank 1 is the model the router sends that request to. expected_cost_usd is the cost of a typical request of that kind at live list prices.

Cite

The Router Index, Nivana (AW3 Technology), 2026-11-20. https://nivana.one/router-index. CC BY 4.0.


Made by Nivana, the AI that sends every request to the best model for it: one chat app and one API, with no model picker. Built by AW3 Technology.

Downloads last month
77