Version 1.0 · 78 business processes × 19 models = 1,482 modelled outcomes · list prices as of 18 Jul 2026
Across every process and model in this index, model choice explains 89.4% of the variation in cost per outcome. The process being automated explains just 10.6%. Put plainly: which model you run matters roughly 8 times more to your bill than which workflow you point it at.
$0.2580Median cost per outcome, all 1,482 cells
69×Between the cheapest and dearest model median
21×Median agentic multiplier vs one chat turn
2.2×Spread across all 16 business categories
Middle 50% of outcomes fall between $0.0803 and $0.4650; the 10th to 90th percentile band runs $0.0420 to $0.8640.
Cost per outcome, by model
Each dot is the median across all 78 processes. Each bar covers the middle 50%. The axis is logarithmic because the range demands it: GPT-4o mini at $0.0227 and Claude Fable 5 at $1.58 are the same unit of work.
Model
Tier
In $/1M
Out $/1M
Median
p25
p75
GPT-4o mini
Efficient
0.15
0.6
$0.0227
$0.0197
$0.0276
Llama 4 Maverick
Balanced
0.27
0.85
$0.0395
$0.0341
$0.0480
Gemini 2.5 Flash
Efficient
0.3
2.5
$0.0533
$0.0473
$0.0646
GPT-4.1 mini
Efficient
0.4
1.6
$0.0606
$0.0526
$0.0736
DeepSeek V4
Balanced
0.435
0.87
$0.0607
$0.0520
$0.0738
Kimi K2.6
Balanced
0.95
4
$0.1451
$0.1261
$0.1762
Claude Haiku 4.5
Efficient
1
5
$0.1575
$0.1375
$0.1912
Grok 4.3
Frontier
1.25
2.5
$0.1744
$0.1494
$0.2120
Gemini 2.5 Pro
Frontier
1.25
10
$0.2194
$0.1944
$0.2660
Mistral Large 3
Frontier
2
6
$0.2910
$0.2510
$0.3536
GPT-4.1
Frontier
2
8
$0.3030
$0.2630
$0.3680
Claude Sonnet 5
Balanced
2
10
$0.3150
$0.2750
$0.3824
GPT-4o
Balanced
2.5
10
$0.3787
$0.3288
$0.4600
GPT-5.4
Frontier
2.5
15
$0.4088
$0.3588
$0.4960
Claude Sonnet 4.6
Balanced
3
15
$0.4725
$0.4125
$0.5736
Kimi K3
Frontier
3
15
$0.4725
$0.4125
$0.5736
Claude Opus 4.8
Frontier
5
25
$0.7875
$0.6875
$0.9560
GPT-5.5
Frontier
5
30
$0.8175
$0.7175
$0.9920
Claude Fable 5
Efficient
10
50
$1.58
$1.38
$1.91
Why the category barely matters
Rows shift dramatically. Columns barely move. Every category lands within a 2.2× band of every other, from Software Engineering / DevOps at $0.3930 down to Supply Chain & Logistics at $0.1800. That is the same finding as the headline, seen from above: the lever is the model, not the department.
The agentic multiplier
One agent-produced outcome is not one model call. Across this index the median outcome costs 21× a single chat turn, with the middle 50% between 18× and 25×, and a maximum of 58×. Published commentary often quotes ranges reaching several hundred times. Those figures describe far heavier configurations than the defaults used here: raise planning loops, tool-result size, retrieval or sub-agents and the multiplier climbs quickly. The 21× figure is what this engine produces at the stated assumptions, and it is the number this index stands behind.
Estimate one process
The charts above are the whole distribution. To price a single process, pick one below. These are typical figures, the median across models in each class; for a specific model and your own assumptions, open the full estimator.
Method
Every cell is one full agent run of one process on one model, priced at published list rates as of 18 Jul 2026. The basket is fixed at 78 processes across 16 categories, modelled as step-level graphs where each step is typed by who does the work: AI agent, human assisted by AI, human only, or system. Only the AI-touching steps consume tokens. Cost is driven by context multiplied by calls, with the running context growing as the agent works.
Fixed assumptions for every cell in this version: base context 4000, prompt tokens/call 800, output tokens/call 600, planning loops 1, tool result tokens 2500, retrieval tokens 1500, sub-agents 1.
Central tendency is reported as the median, never the mean. The distribution is right-skewed, so a mean would be dragged upward by the expensive tail and would describe no real workload.
Two limits worth stating plainly. First, these are modelled estimates from a transparent cost engine, not metered production telemetry. They are built to be checked, not to be believed on authority: the assumptions are listed above and the full dataset is downloadable. Validate against your own traces before committing budget.
Second, the 89/11 split is partly a product of holding assumptions constant across every process. Workflows differ in step count and step type, but they share one assumption set. Tune assumptions per process and workflow variance would rise and the gap would narrow. The direction of the finding is robust, because model list prices genuinely span a 67× range on input alone, but the precise split is a property of this method.
CitationProcessRaven. The Agentic AI Cost Index, v1.0. 18 Jul 2026. https://processraven.tech/cost-index/
Free to cite and reproduce with attribution and a link. If you publish a figure from it, please include the version and the price date, because both move.
Index version 1.0. Generated 2026-07-18. Prices as of 18 Jul 2026. Figures are illustrative estimates, not a quote, forecast or financial advice.
