The short answer
- Most tasks completed
- GPT-5.6 Sol and GPT-5.5
- 42 of 55 tasks (76.4%)
- Best value
- GPT-5.6 Luna
- $0.12 per correct answer, 67.3% of tasks completed
- Fastest
- GPT-5.4
- 39s per task
- Most expensive
- GPT-5.6 Sol
- $2.76 per task — 35× the cheapest, for 9.1 points more tasks completed
No model completed all 55 tasks. The field runs from 76.4% down to 16.4%, so the question is less which model is best than which is good enough for the work in front of you — and what you are prepared to pay per correct answer.
All 17 models, ranked
Ranked by how many of the 55 tasks each completed correctly. Cost per correct answer divides the bill by the tasks that came back right, which is what you actually pay when a failed task has to be redone.
| Model | Provider | Completed | Completion rate | Cost / task | Cost / correct answer | Speed |
|---|---|---|---|---|---|---|
| GPT-5.6 Sol | Azure Foundry | 42/55 | 76.4% | $2.76 | $3.61 | 90s |
| GPT-5.5 | Azure Foundry | 42/55 | 76.4% | $1.18 | $1.54 | 113s |
| Claude Opus 4.6 | Bedrock | 41/55 | 74.5% | $2.27 | $3.04 | 110s |
| GPT-5.6 Terra | Azure Foundry | 40/55 | 72.7% | $1.00 | $1.37 | 74s |
| DeepSeek V4.1 Flash | Fireworks | 38/55 | 69.1% | $0.18 | $0.26 | 209s |
| GPT-5.6 Luna | Azure Foundry | 37/55 | 67.3% | $0.08 | $0.12 | 291s |
| Claude Opus 4.8 | Bedrock | 37/55 | 67.3% | $1.52 | $2.26 | 68s |
| Kimi K3 (FW) | Azure Foundry | 37/55 | 67.3% | $0.31 | $0.46 | 118s |
| Claude Sonnet 5 | Bedrock | 36/55 | 65.5% | $0.86 | $1.32 | 69s |
| DeepSeek V4 Pro | Azure Foundry | 36/55 | 65.5% | $0.55 | $0.83 | 109s |
| Claude Sonnet 4.6 | Bedrock | 36/55 | 65.5% | $0.99 | $1.51 | 83s |
| Claude Opus 5 | Bedrock | 35/55 | 63.6% | $1.64 | $2.58 | 76s |
| DeepSeek V4 Flash | Azure Foundry | 35/55 | 63.6% | $0.09 | $0.14 | 155s |
| GLM 5.3 (FW) | Azure Foundry | 35/55 | 63.6% | $0.30 | $0.47 | 342s |
| GPT-5.4 | Azure Foundry | 29/55 | 52.7% | $0.71 | $1.35 | 39s |
| Claude Haiku 4.5 | Bedrock | 25/55 | 45.5% | $0.26 | $0.57 | 49s |
| Cohere Command A+ | Azure Foundry | 9/55 | 16.4% | $1.31 | $7.98 | 150s |
Run of September 15, 2026. Free to download and reuse with attribution — CSV or JSON, CC BY 4.0.
The best model depends on the task
The overall figure hides the thing that matters most in practice: models that look alike in aggregate diverge sharply by the kind of work. Five different models lead the seven categories.
| Kind of work | Best model | Its score | Field average |
|---|---|---|---|
| Document Ingestion | GPT-5.6 Sol | 5/6 | 42.2% |
| Metric Lookups | GPT-5.6 Sol | 12/12 | 88.7% |
| Financial Statements | DeepSeek V4.1 Flash | 5/7 | 43.7% |
| Error Detection & Repair | Claude Sonnet 5 | 6/7 | 60.5% |
| Dashboards & Widgets | GPT-5.6 Terra | 8/8 | 84.6% |
| Charts | GPT-5.6 Luna | 3/4 | 66.2% |
| Multi-Step Requests | GPT-5.5 | 4/6 | 54.9% |
Questions people ask
Which AI model is most accurate for private-markets financial analysis?
GPT-5.6 Sol and GPT-5.5 tied at the top, each completing 42 of 55 tasks (76.4%). No model passed everything. Run-to-run variation is about ±1.8 tasks, so any gap of two tasks or fewer should be read as a tie rather than a ranking.
Is ChatGPT accurate enough for fund reporting?
Not on its own. The best model tested completed 76.4% of tasks correctly, which means roughly one task in four came back wrong. The failures are rarely obvious — models loaded rows they were told to exclude, or produced figures that look plausible and surface later as unexplained variances. A model is a capable component inside a system that checks its work; it is not a substitute for one.
What does it cost to run an AI agent on a financial task?
Between $0.08 and $2.76 per task across the models tested — a 35× spread. Cost per correct answer is the more useful number, because a failed task has to be redone: that ranges from $0.12 (GPT-5.6 Luna) to $7.98 (Cohere Command A+).
Which AI model gives the best value for financial analysis?
GPT-5.6 Luna, at $0.12 per correct answer. It completed 37 of 55 tasks — 9.1 percentage points behind the leaders, at roughly 35× less per task.
Do more expensive models perform better on financial work?
Barely. GPT-5.6 Sol costs 35× more per task than GPT-5.6 Luna and completes 9.1 percentage points more of the tasks. Price is a poor predictor of how much work a model gets right; the task type matters far more than the price tier.
Which model is best at reading capital-call notices and trial balances?
GPT-5.6 Sol led document ingestion with 5 of 6, and no model passed all six. Ingestion was the hardest category in the benchmark — the field averaged 42.2% — because it requires following instructions exactly across multi-step workflows, not just reading a document.
How were these models tested?
17 models, one environment, one grader — the environment is ours, none of the models are. Every model ran the same 55 tasks inside a production private-markets system, using the same documents, the same tools and the same database, with 935 graded runs in total. Grading is deterministic where a task has one right answer and rubric-based where it does not.
How the test was run
Every model received the same 55 tasks, the same source documents, the same tool definitions and the same database, inside a production private-markets system — not a synthetic test set. 935 graded runs in total. Where a task has a single right answer, grading is deterministic; where it does not, it is scored against a written rubric.
17 models, one environment, one grader. The environment is ours; none of the models are. The models were reached through Azure Foundry, AWS Bedrock and Fireworks on published list pricing.
Run-to-run variation is about ±1.8 tasks at 95% confidence, so a two-task gap is a tie, not a ranking. Latency is end-to-end task time including tool calls, database writes and approval rounds — not raw model inference speed.