RU

Model index for development work

Public leaderboards are half science and trivia. This index counts what I actually give models: code edits, layouts, deploys, short questions. Same independent benchmarks, weighted by my real work.

Updated 20 October 2026 · 22 models · version v2 · methodology

Index

Sonnet 5.5 (max) and Opus 5.5 (max) score highest, followed by Opus 5.5 (xhigh) and Opus 5.5 (high).

Cheapest strong model

Among models scoring 55+, GPT-6.1 Sol (max) is cheapest at $0.72 per task, then Opus 5.5 (medium) at $1.34.

Points per $1

Among models scoring 50+, DeepSeek V4.1 Flash 1017 (max) (198) and GPT-6.1 Sol (max) (79) give the most points per dollar.

Code and design

Opus 5.5 (xhigh) leads at code (56.6), Opus 5.5 (max) at design (65.8).

Highlight
My index

Top 10 · higher is better

63.3
Sonnet 5.5 max
61.9
Opus 5.5 max
61.8
Opus 5.5 xhigh
60.2
Opus 5.5 high
59.3
Sonnet 5.5 xhigh
58.9
Opus 5.5 medium
57.5
GPT-6 Astra
57.1
Fable 5.1
56.6
GPT-6.1 Sol
55.2
Gemini 4 Argon
My index

Top 10 · higher is better
Cost per task

Same models · $ per AA index task · lower is better

$0.72
GPT-6.1 Sol
$1.3
Opus 5.5 medium
$1.8
Opus 5.5 high
$2.0
Gemini 4 Argon
$2.0
Sonnet 5.5 xhigh
$3.3
GPT-6 Astra
$3.5
Opus 5.5 xhigh
$5.5
Sonnet 5.5 max
$6.0
Opus 5.5 max
$7.6
Fable 5.1
Cost per task

Same models · $ per AA index task · lower is better
Points per $1

Same models · index / cost per task · higher is better

79
GPT-6.1 Sol
44
Opus 5.5 medium
33
Opus 5.5 high
30
Sonnet 5.5 xhigh
28
Gemini 4 Argon
18
Opus 5.5 xhigh
18
GPT-6 Astra
12
Sonnet 5.5 max
10
Opus 5.5 max
7
Fable 5.1
Points per $1

Same models · index / cost per task · higher is better

Index

Index on my tasks

Weighted mean of 9 parts · higher is better · “AA index” tab for comparison

63.3
Sonnet 5.5 max
61.9
Opus 5.5 max
61.8
Opus 5.5 xhigh
60.2
Opus 5.5 high
59.3
Sonnet 5.5 xhigh
58.9
Opus 5.5 medium
57.5
GPT-6 Astra
57.1
Fable 5.1
56.6
GPT-6.1 Sol
55.2
Gemini 4 Argon
51.5
DeepSeek V4.1 Flash
49.0
GLM-5.3
48.6
Muse Spark 1.3
47.9
Haiku 5.5 max
47.7
Grok 4.7 xhigh
47.3
Grok 4.7 high
47.2
GLM-5.3-Flash
46.4
Haiku 5.5 xhigh
46.3
Qwen3.8 Max
45.3
MiMo-V2.6-Pro
44.4
Kimi K3
41.2
Gemini 3.8 Flash
58
Opus 5.5 max
56
Sonnet 5.5 max
56
Opus 5.5 xhigh
54
Opus 5.5 high
53
GPT-6 Astra
53
Fable 5.1
53
Gemini 4 Argon
52
Sonnet 5.5 xhigh
52
GPT-6.1 Sol
51
Opus 5.5 medium
48
Muse Spark 1.3
46
Grok 4.7 xhigh
46
Grok 4.7 high
46
MiMo-V2.6-Pro
45
DeepSeek V4.1 Flash
45
GLM-5.3
45
Qwen3.8 Max
44
Kimi K3
43
Haiku 5.5 max
42
GLM-5.3-Flash
41
Haiku 5.5 xhigh
41
Gemini 3.8 Flash

A 1–2 point gap is noise. * — answer style taken from the max version of the same model: AA measures it only there.

Index on my tasks

Weighted mean of 9 parts · higher is better · “AA index” tab for comparison
Index vs cost per task

Cost of one AA index task, log scale · dashed line is the Pareto frontier

best value: above median score, below median cost 40 45 50 55 60 65 $0.1 $0.2 $0.5 $1 $2 $5 $10 $ per task, log scale my index Sonnet 5.5 max Opus 5.5 max Opus 5.5 xhigh Opus 5.5 high Sonnet 5.5 xhigh Opus 5.5 medium GPT-6 Astra Fable 5.1 GPT-6.1 Sol Gemini 4 Argon DeepSeek V4.1 Flash GLM-5.3 Muse Spark 1.3 Haiku 5.5 max Grok 4.7 xhigh Grok 4.7 high GLM-5.3-Flash Haiku 5.5 xhigh Qwen3.8 Max MiMo-V2.6-Pro Kimi K3 Gemini 3.8 Flash

The Pareto frontier: models with nothing cheaper that scores higher. Anything below it is overpaying.

Index vs cost per task

Cost of one AA index task, log scale · dashed line is the Pareto frontier

What the index is made of

Index parts

Score per task type, 0–100 · tab name shows the weight

57
Opus 5.5 xhigh
56
Opus 5.5 max
55
Opus 5.5 high
55
Opus 5.5 medium
55
Sonnet 5.5 max
53
GPT-6 Astra
53
GPT-6.1 Sol
52
Sonnet 5.5 xhigh
51
Fable 5.1
50
Gemini 4 Argon
41
GLM-5.3
41
Qwen3.8 Max
40
MiMo-V2.6-Pro
39
DeepSeek V4.1 Flash
37
Haiku 5.5 xhigh
37
Haiku 5.5 max
37
Muse Spark 1.3
37
GLM-5.3-Flash
30
Grok 4.7 xhigh
30
Grok 4.7 high
28
Kimi K3
28
Gemini 3.8 Flash
66
Opus 5.5 max
64
GPT-6 Astra
63
Sonnet 5.5 max
63
Opus 5.5 xhigh
63
Opus 5.5 high
63
Opus 5.5 medium
59
Sonnet 5.5 xhigh
59
GPT-6.1 Sol
58
Fable 5.1
58
Gemini 4 Argon
57
Muse Spark 1.3
56
Kimi K3
55
DeepSeek V4.1 Flash
51
MiMo-V2.6-Pro
49
Gemini 3.8 Flash
48
Haiku 5.5 max
47
Qwen3.8 Max
47
Haiku 5.5 xhigh
47
Grok 4.7 xhigh
46
Grok 4.7 high
46
GLM-5.3-Flash
44
GLM-5.3
79
Fable 5.1
78
Sonnet 5.5 max
78
Sonnet 5.5 xhigh
76
Opus 5.5 max
76
Opus 5.5 xhigh
76
Opus 5.5 high
76
Opus 5.5 medium
73
DeepSeek V4.1 Flash
69
GPT-6.1 Sol
69
Haiku 5.5 max
69
Haiku 5.5 xhigh
66
GPT-6 Astra
65
GLM-5.3-Flash
65
GLM-5.3
65
Grok 4.7 xhigh
65
Grok 4.7 high
62
Kimi K3
57
Muse Spark 1.3
54
Gemini 3.8 Flash
47
Gemini 4 Argon
40
Qwen3.8 Max
39
MiMo-V2.6-Pro
68
Sonnet 5.5 max
67
Gemini 4 Argon
65
Opus 5.5 max
64
GPT-6 Astra
62
Opus 5.5 xhigh
61
Sonnet 5.5 xhigh
61
GPT-6.1 Sol
60
Opus 5.5 high
57
Opus 5.5 medium
56
Fable 5.1
52
GLM-5.3
50
DeepSeek V4.1 Flash
48
Qwen3.8 Max
47
MiMo-V2.6-Pro
47
GLM-5.3-Flash
46
Grok 4.7 xhigh
46
Muse Spark 1.3
44
Grok 4.7 high
40
Gemini 3.8 Flash
35
Kimi K3
34
Haiku 5.5 max
33
Haiku 5.5 xhigh
64
Gemini 4 Argon
57
Qwen3.8 Max
57
GLM-5.3-Flash
55
GLM-5.3
54
Muse Spark 1.3
52
MiMo-V2.6-Pro
52
Grok 4.7 xhigh
51
Grok 4.7 high
51
Haiku 5.5 max
51
Haiku 5.5 xhigh
49
Sonnet 5.5 max
48
GPT-6 Astra
47
GPT-6.1 Sol
47
Opus 5.5 max
45
Kimi K3
44
Opus 5.5 medium
44
Opus 5.5 xhigh
43
Opus 5.5 high
42
Sonnet 5.5 xhigh
41
Gemini 3.8 Flash
39
DeepSeek V4.1 Flash
38
Fable 5.1
58
Sonnet 5.5 max
58
Fable 5.1
56
Opus 5.5 max
55
Grok 4.7 high
54
Haiku 5.5 max
53
Grok 4.7 xhigh
51
GPT-6 Astra
50
GPT-6.1 Sol
49
Gemini 4 Argon
46
Gemini 3.8 Flash
45
Qwen3.8 Max
39
GLM-5.3
39
Kimi K3
38
Muse Spark 1.3
38
MiMo-V2.6-Pro
25
GLM-5.3-Flash
79
Opus 5.5 max
79
Sonnet 5.5 max
76
Opus 5.5 xhigh
69
Sonnet 5.5 xhigh
66
Fable 5.1
63
Opus 5.5 high
61
Grok 4.7 xhigh
60
Grok 4.7 high
58
DeepSeek V4.1 Flash
56
Qwen3.8 Max
54
Muse Spark 1.3
51
Opus 5.5 medium
50
MiMo-V2.6-Pro
50
Haiku 5.5 max
47
GLM-5.3
47
GPT-6.1 Sol
46
GPT-6 Astra
44
Gemini 4 Argon
43
GLM-5.3-Flash
39
Haiku 5.5 xhigh
38
Kimi K3
19
Gemini 3.8 Flash
53
Gemini 3.8 Flash
51
GLM-5.3-Flash
50
Fable 5.1
49
GPT-6 Astra
48
Kimi K3
46
GLM-5.3
42
Grok 4.7 xhigh
40
Qwen3.8 Max
38
Opus 5.5 max
33
Muse Spark 1.3
81
Gemini 4 Argon
68
Grok 4.7 xhigh
66
GLM-5.3-Flash
66
GLM-5.3
66
Grok 4.7 high
64
Qwen3.8 Max
63
Muse Spark 1.3
62
Sonnet 5.5 max
59
MiMo-V2.6-Pro
59
GPT-6 Astra
56
Opus 5.5 max
55
GPT-6.1 Sol
53
Kimi K3
52
Gemini 3.8 Flash
51
Sonnet 5.5 xhigh
50
DeepSeek V4.1 Flash
50
Opus 5.5 xhigh
48
Opus 5.5 high
48
Haiku 5.5 max
46
Opus 5.5 medium
46
Haiku 5.5 xhigh
43
Fable 5.1
Index parts

Score per task type, 0–100 · tab name shows the weight
Benchmarks

Raw independent benchmark results the parts are built from

code edits in real repositories · percent solved

57.1
Opus 5.5 medium
54.0
Opus 5.5 high
53.5
Opus 5.5 xhigh
52.2
Opus 5.5 max
49.1
GPT-6.1 Sol
49.0
Fable 5.1
47.0
GPT-6 Astra
46.1
Sonnet 5.5 xhigh
45.8
Sonnet 5.5 max
45.5
Haiku 5.5 xhigh
45.4
MiMo-V2.6-Pro
43.7
DeepSeek V4.1 Flash
43.6
Kimi K3
43.5
Gemini 4 Argon
43.2
Qwen3.8 Max
41.8
Haiku 5.5 max
41.0
GLM-5.3-Flash
40.6
Muse Spark 1.3
40.1
GLM-5.3
36.4
Gemini 3.8 Flash
34.2
Grok 4.7 high
33.8
Grok 4.7 xhigh

terminal work: builds, deploys · percent solved

63.6
Sonnet 5.5 max
59.6
Opus 5.5 max
59.6
Opus 5.5 xhigh
59.1
GPT-6 Astra
57.1
Sonnet 5.5 xhigh
57.1
Gemini 4 Argon
56.6
Opus 5.5 high
56.1
GPT-6.1 Sol
52.5
Opus 5.5 medium
52.0
Fable 5.1
41.9
GLM-5.3
38.9
Qwen3.8 Max
34.8
MiMo-V2.6-Pro
33.8
DeepSeek V4.1 Flash
33.3
Muse Spark 1.3
32.8
Haiku 5.5 max
32.8
GLM-5.3-Flash
29.3
Haiku 5.5 xhigh
25.8
Grok 4.7 xhigh
24.7
Grok 4.7 high
19.7
Gemini 3.8 Flash
12.6
Kimi K3

UI layouts, human votes · Elo rating

1,390
Opus 5.5 max
1,390
Opus 5.5 xhigh
1,390
Opus 5.5 high
1,390
Opus 5.5 medium
1,379
GPT-6 Astra
1,365
Sonnet 5.5 max
1,365
Sonnet 5.5 xhigh
1,361
Kimi K3
1,352
Muse Spark 1.3
1,346
GPT-6.1 Sol
1,336
Fable 5.1
1,333
DeepSeek V4.1 Flash
1,328
Gemini 4 Argon
1,317
MiMo-V2.6-Pro
1,304
GLM-5.3
1,304
Gemini 3.8 Flash
1,293
Qwen3.8 Max
1,282
GLM-5.3-Flash
1,279
Haiku 5.5 max
1,279
Haiku 5.5 xhigh
1,271
Grok 4.7 xhigh
1,271
Grok 4.7 high

multi-step tool use · percent solved

77.5
Gemini 4 Argon
71.8
Sonnet 5.5 max
69.5
Opus 5.5 max
68.5
GPT-6 Astra
66.6
DeepSeek V4.1 Flash
65.6
Grok 4.7 xhigh
65.5
Sonnet 5.5 xhigh
65.0
Opus 5.5 xhigh
64.9
GPT-6.1 Sol
63.5
Grok 4.7 high
63.2
Opus 5.5 high
62.2
GLM-5.3
61.2
Opus 5.5 medium
60.4
GLM-5.3-Flash
59.9
Gemini 3.8 Flash
59.4
Fable 5.1
58.6
MiMo-V2.6-Pro
58.3
Kimi K3
57.9
Muse Spark 1.3
56.2
Qwen3.8 Max
36.0
Haiku 5.5 xhigh
35.4
Haiku 5.5 max

does not invent facts · percent solved

84.9
Gemini 4 Argon
72.4
GLM-5.3-Flash
71.2
Qwen3.8 Max
70.7
Grok 4.7 xhigh
70.4
GLM-5.3
67.6
Grok 4.7 high
67.1
Muse Spark 1.3
59.6
Haiku 5.5 max
59.4
MiMo-V2.6-Pro
55.6
Haiku 5.5 xhigh
53.0
Sonnet 5.5 max
48.7
GPT-6 Astra
46.8
Kimi K3
45.7
GPT-6.1 Sol
44.8
Gemini 3.8 Flash
41.4
Opus 5.5 max
37.1
Sonnet 5.5 xhigh
34.3
Opus 5.5 xhigh
33.2
DeepSeek V4.1 Flash
32.4
Opus 5.5 high
31.6
Opus 5.5 medium
27.4
Fable 5.1

data analysis · percent solved

57.5
Sonnet 5.5 max
57.5
Fable 5.1
56.2
Opus 5.5 max
55.0
Grok 4.7 high
53.8
Haiku 5.5 max
52.5
Grok 4.7 xhigh
51.2
GPT-6 Astra
50.0
GPT-6.1 Sol
48.8
Gemini 4 Argon
46.2
Gemini 3.8 Flash
45.0
Qwen3.8 Max
38.8
GLM-5.3
38.8
Kimi K3
37.5
Muse Spark 1.3
37.5
MiMo-V2.6-Pro
25.0
GLM-5.3-Flash

understanding a loose brief · Elo rating

1,817
Sonnet 5.5 max
1,803
Opus 5.5 max
1,764
Opus 5.5 xhigh
1,751
Sonnet 5.5 xhigh
1,686
Opus 5.5 high
1,672
Fable 5.1
1,642
Grok 4.7 xhigh
1,633
Grok 4.7 high
1,625
Opus 5.5 medium
1,613
Qwen3.8 Max
1,581
Muse Spark 1.3
1,577
Haiku 5.5 max
1,570
DeepSeek V4.1 Flash
1,569
GPT-6 Astra
1,557
GPT-6.1 Sol
1,532
Haiku 5.5 xhigh
1,515
MiMo-V2.6-Pro
1,507
GLM-5.3
1,499
Kimi K3
1,486
Gemini 4 Argon
1,453
GLM-5.3-Flash
1,203
Gemini 3.8 Flash

real work tasks · Elo rating

1,866
Opus 5.5 max
1,838
Sonnet 5.5 max
1,831
Opus 5.5 xhigh
1,754
Fable 5.1
1,746
DeepSeek V4.1 Flash
1,728
Sonnet 5.5 xhigh
1,711
Grok 4.7 xhigh
1,706
Grok 4.7 high
1,704
Opus 5.5 high
1,684
MiMo-V2.6-Pro
1,681
Muse Spark 1.3
1,665
Qwen3.8 Max
1,650
GLM-5.3
1,642
GLM-5.3-Flash
1,623
Gemini 4 Argon
1,618
Haiku 5.5 max
1,594
GPT-6.1 Sol
1,584
Opus 5.5 medium
1,573
GPT-6 Astra
1,533
Kimi K3
1,513
Haiku 5.5 xhigh
1,436
Gemini 3.8 Flash

finding causes of incidents · percent solved

52.5
Gemini 3.8 Flash
51.2
GLM-5.3-Flash
49.5
Fable 5.1
48.6
GPT-6 Astra
47.7
Kimi K3
46.1
GLM-5.3
42.1
Grok 4.7 xhigh
40.3
Qwen3.8 Max
38.2
Opus 5.5 max
33.2
Muse Spark 1.3

3D scenes from a brief · Elo rating

1,197
Sonnet 5.5 max
1,191
Opus 5.5 max
1,166
GPT-6 Astra
1,154
Gemini 4 Argon
1,133
Fable 5.1
1,118
GPT-6.1 Sol
1,037
Muse Spark 1.3
997
Haiku 5.5 max
995
Grok 4.7 xhigh
981
Kimi K3
980
MiMo-V2.6-Pro
969
Gemini 3.8 Flash
948
Qwen3.8 Max
935
GLM-5.3-Flash
787
GLM-5.3
Benchmarks

Raw independent benchmark results the parts are built from

Weights

Task type weight

Weight is the mean of three shares: tasks, spend and corrections · tracker 2026-09-24…2026-10-18, 5,107 tasks

  • share of tasks
  • share of spend
  • share of corrections
  • weight in index
20.4
39.9
12.5
24.3
Code
12.9
19.8
29.0
20.6
Design
19.1
10.5
17.6
15.7
Short questions
10.9
6.3
13.3
10.2
CI and deploy
14.1
11.3
1.6
9.0
Review
6.7
3.3
9.0
6.3
Data and SQL
6.2
2.1
10.6
6.3
Understand the ask
5.0
2.7
3.9
3.9
Find the problem
4.8
4.0
2.4
3.7
Web research

Corrections do not grade a model (nearly all are on Opus 5.5); they show where tasks are hard and need a stricter benchmark.

Task type weight

Weight is the mean of three shares: tasks, spend and corrections · tracker 2026-09-24…2026-10-18, 5,107 tasks
Benchmark Used for Weight
Terminal-Bench 4.0 code ½, CI ½ 17.3%
FrontierCode 1.1 code ½, review ½ 16.7%
AA-PersonalityBench short questions: style and brevity 15.7%
Design Arena design ¾ 15.5%
AutomationBench-AA CI ½, web ½ 6.9%
AA-Omniscience review ½, web ½ 6.4%
AA-AnalystAgent data 6.3%
AA-3D design ¼ 5.2%
ITBench-AA find the problem 3.9%
AA-Briefcase understand ½ 3.2%
GDPval-AA understand ½ 3.2%

Leaderboard

All models

Click a header to sort
Sonnet 5.5 (max) Anthropic63.35655637868495879—62$5.4612
Opus 5.5 (max) Anthropic61.958566676654756793856$5.9810
Opus 5.5 (xhigh) * Anthropic61.8565763766244—76—50$3.4618
Opus 5.5 (high) * Anthropic60.2545563766043—63—48$1.8233
Sonnet 5.5 (xhigh) * Anthropic59.3525259786142—69—51$2.0130
Opus 5.5 (medium) * Anthropic58.9515563765744—51—46$1.3444
GPT-6 Astra (max) OpenAI57.553536466644851464959$3.2618
Fable 5.1 (max) Anthropic57.153515879563858665043$7.637
GPT-6.1 Sol (max) OpenAI56.65253596961475047—55$0.7279
Gemini 4 Argon (high) Google55.25350584767644944—81$1.9928
DeepSeek V4.1 Flash 1017 (max) DeepSeek51.5453955735039—58—50$0.26198
GLM-5.3 (max) Z AI49.045414465525539474666$2.0124
Muse Spark 1.3 (max) Meta48.648375757465438543363$1.6030
Haiku 5.5 (max) Anthropic47.94337486934515450—48$0.21228
Grok 4.7 (xhigh) SpaceXAI47.746304765465253614268$3.7413
Grok 4.7 (high) * SpaceXAI47.34630466544515560—66$2.7317
GLM-5.3-Flash Z AI47.242374665475725435166$0.28169
Haiku 5.5 (xhigh) * Anthropic46.4413747693351—39—46$0.12387
Qwen3.8 Max (0902) Alibaba46.345414740485745564064$5.419
MiMo-V2.6-Pro Xiaomi45.34640513947523850—59$0.13348
Kimi K3 (max) Kimi44.444285662354539384853$2.0022
Gemini 3.8 Flash (high) Google41.241284954404146195352$1.2433
All models

Click a header to sort

Methodology

  • Tasks come from my Claude Code logs for 2026-09-24…2026-10-18: 5,107 tasks in 9 types. Customer support is excluded.
  • A type’s weight is the mean of three shares: how many tasks, how much money and how many corrections.
  • Each type maps to the closest independent benchmark (Artificial Analysis, Design Arena). Vendor numbers are never used. Elo becomes a win chance: 100 / (1 + 10^((base − Elo) / 400)).
  • If a model lacks a benchmark, the other weights are renormalised; only models with at least 85% of the weight measured are ranked.
  • Sources: Artificial Analysis (v4.3.2), Design Arena (2026-10-20 18:07 UTC).

Version history

  1. 20 October 2026

    v2: 9 task types weighted by the mean of task, spend and correction shares; design via Design Arena; 22 models from all labs.

  2. 14 October 2026

    v1: agentic coding 45%, understanding 45%, not hallucinating 10% — five benchmarks, Claude only.

AI spend and subscriptions are on a separate page: how much AI coding fits in a subscription.