Which AI can actually raise money?

We put 18 AI models through 38 tasks your development team handles every week: appeals, donor reports, grant proposals, major gift strategy. We judged each appeal against a control letter that has already proven it raises money. Everything else was graded with expert rubrics and answer keys. Then a working fundraiser checked every score.

v3.1Benchmark
18models tested
38fundraising tests
14weighted categories
114graded runs per model

Best AI fundraising LLM

Weighted score, 0 to 100
  1. 1Claude Fable 5.2Anthropic89.96
  2. 2GPT-6.2 AstraOpenAI89.59
  3. 3Claude Opus 5.5 HighAnthropic87.72

Last update: November 21, 2026

Good with data. Still learning to ask.

76.9

Appeals are every model's weakest category

The best appeal score was 76.9, from GPT-6.2 Astra. Appeals carry 18% of the total, more than any other category, and all 18 models scored lower there than anywhere else.

97.6

Reporting and testing are close to solved

The top models scored in the mid 90s on reporting, testing and optimization, prospect research and ethics. GPT-6.2 Astra's 97.6 on testing was the highest single score.

65.9

Small models keep the voice, lose the math

Claude Haiku 5.5 scored 84.5 on storytelling but 65.9 on donor data analysis. If your team runs reports through AI, the model size matters.

The rankings

Rank by the weighted overall score, or pick one category to see who is best at the work your team does most.

Rank
Model
Overall score
Overall
Appeals
Released
1
Claude Fable 5.2
Anthropic
89.96
89.96
76.2
Nov 9, 2026
2
GPT-6.2 Astra
OpenAI
89.59
89.59
76.9
Nov 17, 2026
3
Claude Opus 5.5 High
Anthropic
87.72
87.72
74.3
Sep 22, 2026
4
Gemini 4 Carbon
Google
86.65
86.65
71.3
Oct 28, 2026
5
Gemini 4 Neon
Google
85.38
85.38
69.7
Nov 18, 2026
6
Grok 4.7 Heavy
xAI
84.85
84.85
68.7
Sep 21, 2026
7
Claude Sonnet 5.5
Anthropic
84.25
84.25
70.4
Sep 28, 2026
8
GPT-6.2 Sol
OpenAI
82.60
82.60
66.8
Nov 17, 2026
9
GPT-6 Astra High (Codex)
OpenAI
82.56
82.56
62.3
Sep 3, 2026
10
Qwen4-Max
Alibaba
77.22
77.22
58.2
Nov 16, 2026
11
Mistral Large 4 (High)
Mistral AI
76.18
76.18
60.3
Oct 6, 2026
12
Kimi K3.1 Pro
Moonshot AI
75.76
75.76
56.7
Oct 23, 2026
13
Muse Wonder
Meta
75.51
75.51
59.9
Oct 12, 2026
14
Mistral Think mode (Vibe)
Mistral AI
73.46
73.46
58.7
15
DeepSeek-V4.1-Flash (thinking)
DeepSeek
72.57
72.57
53.0
Sep 10, 2026
16
Claude Haiku 5.5
Anthropic
71.90
71.90
54.4
Oct 7, 2026
17
Gemini 3.8 Flash
Google
67.10
67.10
47.3
Sep 2, 2026
18
Muse Ember
Meta
56.54
56.54
33.4
Nov 9, 2026

Showing 18 of 18 models · Each score combines three graded runs per test · Kit v3.1

Where each model is strong

Every model's score in all 14 categories. The percentage under each code is that category's share of the overall score.

Score90 and up80 to 89.970 to 79.960 to 69.9Below 60
Model
All
AP18%
MG10%
DA10%
ST8%
RP8%
GR7%
PG7%
SP6%
PR5%
TO5%
SC5%
JE5%
DG3%
EV3%
Claude Fable 5.2
90.0
76.2
84.5
88.4
94.0
96.7
95.6
94.2
92.4
96.4
96.9
95.2
97.1
93.5
92.8
GPT-6.2 Astra
89.6
76.9
83.4
89.9
92.4
97.4
94.2
92.5
90.9
95.6
97.6
93.3
94.8
95.0
93.3
Claude Opus 5.5 High
87.7
74.3
82.6
79.8
93.4
95.1
93.8
91.2
89.4
94.7
96.2
95.8
96.3
93.1
92.4
Gemini 4 Carbon
86.7
71.3
79.9
86.7
90.1
95.7
92.4
90.8
88.0
96.2
94.8
90.1
93.9
91.4
89.8
Gemini 4 Neon
85.4
69.7
78.2
84.8
89.4
94.7
91.4
89.2
86.5
94.2
94.2
89.1
94.4
90.4
88.8
Grok 4.7 Heavy
84.8
68.7
77.6
86.2
87.5
94.1
91.7
90.0
88.8
92.8
94.6
87.4
90.4
88.5
87.2
Claude Sonnet 5.5
84.2
70.4
78.6
75.2
91.1
92.2
90.7
86.5
85.4
91.2
93.1
93.4
94.2
90.2
90.1
GPT-6.2 Sol
82.6
66.8
75.5
80.3
88.0
91.7
88.7
84.8
83.8
89.8
91.7
89.4
92.1
89.3
87.4
GPT-6 Astra High (Codex)
82.6
62.3
74.7
87.7
83.4
95.2
89.8
88.1
84.9
92.1
95.7
83.9
91.0
87.0
84.3
Qwen4-Max
77.2
58.2
69.1
79.5
80.7
89.6
84.6
80.1
79.6
85.0
88.8
81.2
84.9
83.9
81.4
Mistral Large 4 (High)
76.2
60.3
74.1
76.0
78.3
87.0
81.8
76.8
76.1
82.7
86.2
79.7
83.9
81.1
79.0
Kimi K3.1 Pro
75.8
56.7
67.2
78.1
79.0
88.5
82.9
77.9
77.9
85.6
87.3
80.7
83.3
82.1
79.5
Muse Wonder
75.5
59.9
66.5
72.8
82.4
85.3
81.3
75.8
76.9
81.6
85.4
83.8
84.4
84.3
82.8
Mistral Think mode (Vibe)
73.5
58.7
72.6
70.4
77.2
84.0
78.7
74.0
72.8
79.8
84.0
77.7
79.2
77.6
75.9
DeepSeek-V4.1-Flash (thinking)
72.6
53.0
63.6
76.4
75.7
86.3
79.8
74.8
74.3
80.8
85.8
76.6
79.8
78.2
76.4
Claude Haiku 5.5
71.9
54.4
62.0
65.9
81.9
82.8
79.6
70.4
71.7
76.4
83.1
84.5
85.8
81.2
80.5
Gemini 3.8 Flash
67.1
47.3
56.0
68.1
72.6
82.0
74.4
66.2
68.4
78.1
81.2
73.4
75.2
76.5
72.3
Muse Ember
56.5
33.4
44.9
57.3
64.4
74.9
63.7
53.8
56.9
67.3
72.9
65.3
66.0
67.5
64.1

Category leaders

AP · 18% of score

Appeals and solicitation

GPT-6.2 Astra76.9
Next: Claude Fable 5.276.2
MG · 10% of score

Major gifts and relationship fundraising

Claude Fable 5.284.5
Next: GPT-6.2 Astra83.4
DA · 10% of score

Donor data analysis

GPT-6.2 Astra89.9
Next: Claude Fable 5.288.4
ST · 8% of score

Stewardship and donor care

Claude Fable 5.294.0
Next: Claude Opus 5.5 High93.4
RP · 8% of score

Reporting and dashboards

GPT-6.2 Astra97.4
Next: Claude Fable 5.296.7
GR · 7% of score

Grants and institutional giving

Claude Fable 5.295.6
Next: GPT-6.2 Astra94.2
PG · 7% of score

Planned and complex gifts

Claude Fable 5.294.2
Next: GPT-6.2 Astra92.5
SP · 6% of score

Strategy, planning, and leadership

Claude Fable 5.292.4
Next: GPT-6.2 Astra90.9
PR · 5% of score

Prospect research

Claude Fable 5.296.4
Next: Gemini 4 Carbon96.2
TO · 5% of score

Testing and optimization

GPT-6.2 Astra97.6
Next: Claude Fable 5.296.9
SC · 5% of score

Storytelling and case for support

Claude Opus 5.5 High95.8
Next: Claude Fable 5.295.2
JE · 5% of score

Judgment, ethics, and compliance

Claude Fable 5.297.1
Next: Claude Opus 5.5 High96.3
DG · 3% of score

Digital and multichannel

GPT-6.2 Astra95.0
Next: Claude Fable 5.293.5
EV · 3% of score

Events

GPT-6.2 Astra93.3
Next: Claude Fable 5.292.8

Our methodology

We tested real fundraising tasks in ways AI can execute them.

The test organization
Kindred Fields International

A fictional secular relief and development nonprofit based in Denver, built so no model has seen it before.

$12.2Mannual revenue
12development staff
10,000donors, five years
1

38 tests in 14 categories

Appeals, major gifts, donor data, grants, planned giving, stewardship and more. Each prompt is written the way a senior fundraiser briefs a capable colleague.

2

Three runs per test

Every model runs every test three times, 114 graded runs in all, so one lucky answer cannot carry a score.

3

Graded against a proven control

Appeals are judged by six graders against a control appeal. Rubric tests take the median of three graders. Data tests are checked against an answer key. A human confirms every score.

4

Making things up fails the test

A new fact, figure, name or quote with no basis in the source material is an automatic zero for that run.

How the overall score is weighted

Appeals count most, at 18%, because they're the most widely used and move the most money from everyday donors.

  • Appeals and solicitation18%
  • Major gifts and relationship fundraising10%
  • Donor data analysis10%
  • Stewardship and donor care8%
  • Reporting and dashboards8%
  • Grants and institutional giving7%
  • Planned and complex gifts7%
  • Strategy, planning, and leadership6%
  • Prospect research5%
  • Testing and optimization5%
  • Storytelling and case for support5%
  • Judgment, ethics, and compliance5%
  • Digital and multichannel3%
  • Events3%