How 17 AI Models Perform on Real Private-Markets Finance Work
Last updated: September 15, 2026
We ran 17 frontier models through the same 55 financial tasks — capital-call notices, trial balances, fund statements, dashboards — inside a production system, and graded every run. Here is what each model got right, what it cost, and how long it took.
17 models, one environment, one grader. The environment is ours; none of the models are.
Every model ran the same 55 tasks inside a real fund-reporting system — reading the same documents, using the same tools, writing to the same database — and every run was graded the same way. 935 graded runs in total. Click any model for its own page.
Highest Pass Rate
76.4%
GPT-5.5 (Azure Foundry)
Lowest Cost
$0.08
GPT-5.6 Luna (Azure Foundry)
Fastest Latency
39s
GPT-5.4 (Azure Foundry)
Task Categories
7
17 models · Ingest, Lookups, Statements, Errors, Boards, Charts, Compound
All Results — Overview
All 17 models, ranked by how many of the 55 tasks they completed correctly. Real files, real approval steps, real clarifying questions — not a synthetic test set. Weighted by each model's tested sample count.
Ladder
Two engines share the top of the ladder; none clears the whole suite.
GPT-5.6 Sol and GPT-5.5 both pass 42 of 55 tasks. Cohere Command A+ sits at the bottom on 9. No single engine sweeps every task category.
76.4% → 16.4%
Pass rate, top to bottom
Frontier Result
GPT-5.5 matches GPT-5.6 Sol at a third of the price.
Both engines pass 42 of 55. GPT-5.6 Sol runs $2.76 a task against GPT-5.5's $1.18, and the judge separates them by a tenth of a point.
42/55 each
GPT-5.6 Sol and GPT-5.5, tied on pass rate
Cost Reality
The priciest engine costs 34x the cheapest, for 9 points of pass rate.
GPT-5.6 Sol runs $2.76 per task; GPT-5.6 Luna runs $0.08 and still clears 37 of 55. The gap in tasks completed is a fraction of the gap in cost.
34.5x
Cost spread, cheapest to priciest engine
Overall Pass Rate
Overall Latency
Overall Cost vs Pass Rate
Cost
What it actually costs to put each engine through the work. The left chart is one full pass of all 55 tasks at list API prices; the right divides that bill by the tasks the engine got right, which is the number that matters when a failed task has to be re-run or fixed by hand.
Cost — One Full Pass (55 tasks)
Cost per Passed Task
Latency & Effort
How long the work takes and how much machinery it burns getting there. Full-pass figures are the engine's mean across all 55 tasks scaled to the whole suite, so they show what an unattended overnight run looks like. Tool calls are the clearest proxy for wasted effort — an engine making twice the calls for the same pass rate is thrashing.
Latency — One Full Pass (55 tasks)
Latency — Average per Task
Tool Calls — One Full Pass (55 tasks)
Tool Calls — Average per Task
Full Results Table
All 17 models ranked by overall pass rate across every task category, with tool-call efficiency and cost breakdown. Run-to-run noise is about ±1.8 tasks, so gaps of two tasks or fewer are ties. Click any model — here or on any chart above — for its own page.
| Rank | Model | Provider | Tier | Passed | Pass Rate | Cost/Task | Cost/Suite | Latency | Tool Calls |
|---|---|---|---|---|---|---|---|---|---|
| 1 | GPT-5.6 Sol | Azure Foundry (OpenAI) | High | 42/55 | 76.4% | $2.76 | $152 | 90s | 28.1 |
| 2 | GPT-5.5 | Azure Foundry (OpenAI) | High | 42/55 | 76.4% | $1.18 | $65 | 113s | 27.2 |
| 3 | Claude Opus 4.6 | Bedrock (Anthropic) | High | 41/55 | 74.5% | $2.27 | $125 | 110s | 23.1 |
| 4 | GPT-5.6 Terra | Azure Foundry (OpenAI) | Med | 40/55 | 72.7% | $1.00 | $55 | 74s | 23.9 |
| 5 | DeepSeek V4.1 Flash | Fireworks (direct) | Low | 38/55 | 69.1% | $0.18 | $10 | 209s | 37.3 |
| 6 | GPT-5.6 Luna | Azure Foundry (OpenAI) | Low | 37/55 | 67.3% | $0.08 | $4 | 291s | 22.5 |
| 7 | Claude Opus 4.8 | Bedrock (Anthropic) | High | 37/55 | 67.3% | $1.52 | $84 | 68s | 22.0 |
| 8 | Kimi K3 (FW) | Azure Foundry (Fireworks) | Med | 37/55 | 67.3% | $0.31 | $17 | 118s | 23.6 |
| 9 | Claude Sonnet 5 | Bedrock (Anthropic) | Med | 36/55 | 65.5% | $0.86 | $48 | 69s | 22.7 |
| 10 | DeepSeek V4 Pro | Azure Foundry (DeepSeek) | High | 36/55 | 65.5% | $0.55 | $30 | 109s | 29.8 |
| 11 | Claude Sonnet 4.6 | Bedrock (Anthropic) | Med | 36/55 | 65.5% | $0.99 | $54 | 83s | 23.8 |
| 12 | Claude Opus 5 | Bedrock (Anthropic) | High | 35/55 | 63.6% | $1.64 | $90 | 76s | 23.7 |
| 13 | DeepSeek V4 Flash | Azure Foundry (DeepSeek) | Low | 35/55 | 63.6% | $0.09 | $5 | 155s | 26.6 |
| 14 | GLM 5.3 (FW) | Azure Foundry (Fireworks) | Med | 35/55 | 63.6% | $0.30 | $17 | 342s | 30.0 |
| 15 | GPT-5.4 | Azure Foundry (OpenAI) | Med | 29/55 | 52.7% | $0.71 | $39 | 39s | 18.2 |
| 16 | Claude Haiku 4.5 | Bedrock (Anthropic) | Low | 25/55 | 45.5% | $0.26 | $14 | 49s | 18.7 |
| 17 | Cohere Command A+ | Azure Foundry (Cohere) | Med | 9/55 | 16.4% | $1.31 | $72 | 150s | 37.1 |
Methodology
Task Breakdown
Each model is evaluated across 7 distinct task categories that mirror real product surfaces: ingest harness (document parsing), lookups & metrics, statement generation, error recovery, boards & widgets, chart generation, and compound multi-turn tasks. Pass rate and cost are reported per category so tool-selection tradeoffs are visible per workflow, not just in aggregate.
Scoring
Ingest tasks use deterministic pass/fail checks. Lookups, statements, error-recovery, boards, charts, and compound tasks use rubric-based grading with partial credit, rolled up into a per-category pass count. Sample counts vary slightly by category and model based on task applicability.
Cost & Tool Calls
Cost per task reflects actual token usage at each provider's published pricing for that category's harness. Tool calls counts the average number of tool invocations per task — a proxy for how many round-trips a model needs to complete a workflow, independent of whether it ultimately passes.
Environment
All models are accessed through their respective cloud APIs (AWS Bedrock, Azure Foundry). Each model receives the same tool definitions, system prompt, and data payload within a given category.
See how HenonMAX performs on your workflows
The full methodology is published on this page and the results are free to download. Schedule a demonstration to see the same harness run against your own data.