Eigen Networks · Benchmark findings · October 2026

Two desks on small models, against a frontier loop

An insurance motor claims desk and a bank's KYC desk, each 1,000 cases a day. Arm A: one frontier model in a single agent loop, at published list prices. Arm B: Eigen N with every role on small models inside your perimeter, priced as hardware. Tool calls per case come from the two benchmarks; the anchor run is measured; the desks scale from it.

The numbers

Measured · one run · claim C7 · 27 September 2026

solved
partial coverage, 1,350 EUR, decision written, disposition submitted
34
small model calls (Qwen 3.5 9B, Gemma 4 A4B), plus one structurer call
8
tool calls, right order, none repeated, no repair turn
46,010
tokens inside the perimeter: 37,787 in, 8,223 out
102 s
execution; 300 s with structuring and planning
0.85
grader quality; the clause citation scored 0 on a formatting defect fixed the same night

This run anchors every per case figure on this page: 8 tool calls, 46,010 tokens, 110 card seconds. Other workflows scale with their tool calls. The structurer call in this run still went to a hosted model; every configuration below runs it on a small model inside the perimeter.

Assumed · the two desks

What runs every day

Tool calls per workflow are read from the benchmark case designs and anchored on the measured C7. Runs a day are the assumption for a mid size German insurer and a mid size bank; change them and every number below moves in proportion.

Assumed · where the tokens go

One run, two ways

Frontier loop model per workflow: tool calls plus four turns (three reasoning, one decision), a 2,900 token prompt, an 800 token receipt and a 150 token reply per turn, the whole history sent again every turn, 300 reasoning tokens per turn billed as output. The history sent again is priced at the providers' cached rate, one tenth of list.

Estimated · cost per case

Frontier loop at list prices, Eigen N as hardware

Frontier bars are tokens only. Eigen N bars are the full monthly cost of the server (purchase over 36 months, power, hosting) spread over the desk's 30,000 cases a month. Eigen N has no token price, so the bar is the same for every model pool; pools differ only in how many cases a day one server handles.

Estimated · the whole desk per month

What each desk costs a month

Both desks at 1,000 cases a day. The frontier figures grow with every case; the Eigen N figure is one server until the desk outgrows it.

Estimated · capacity

From one card to the full pool

The same desks on the model pools we run today. The server costs the same whichever pool it serves; what differs is machine time per case, and so how many cases a day one pool handles in a 20 hour window.

The same desks on seven GPU classes

Servers hold two cards, or one 48 or 96 GB card. Both desks fit on one server of any class at 1,000 cases a day except the L4 class, which fits only the claims desk. Latency per case stays at two to five minutes; a desk is a throughput question, and that is what is priced here.

Estimated · break even

Monthly cost as a desk grows from 200 to 2,000 cases a day

Frontier lines are token cost only. Eigen N lines include hardware, power, hosting and the small structurer. Steps are additional servers. Copilot Studio lies above every frontier line.

Months until an owned server has paid for itself, at 1,000 cases a day

Purchase price divided by the monthly saving against the loop in the column. Green within six months, amber within eighteen, grey longer.

Estimated · energy and data

What the desk draws, and what leaves the building

281 kWh
a month for the whole desk on the baseline server, at 60 percent card utilisation
9 Wh
per case, the server's whole draw spread over 30,000 cases; cooling apart
0
bytes of case data to an external model service: planning, Agents, tools and records all inside your perimeter
not disclosed
energy per request of the frontier providers; no number is invented for it

In the product

What a single loop does not give

  • ● in the product today
  • ◐ next
  • ◔ partly, with code you write around the model
  • ○ not there

From one case to a company brain

The networks the desks leave behind

Every case the planner solves is recorded as a network: the plan, the Agents, the tools each one called, the evidence each one cited and the verdicts. Claim C1 leaves a four step network. Claim C2, the same claim with the invoice missing, lands on the same approved plan with one additional branch: request the invoice, then continue. The two share three of their four steps.

The KYC desk builds its own networks in its own domain, with its own access boundary. Nothing crosses from claims to KYC unless a person approves it. Over months the approved networks of a domain become the institution's working knowledge of how its cases are handled: searchable, versioned, with the evidence attached. That is EigenBrain, and Nexus is the knowledge system that holds it.

Next step

The pilot: one workflow, your cases, your reviewers

Eight weeks on one agreed workflow inside your environment, on 30 to 50 historical cases, read only, nothing changes in live records. Your reviewers judge every prepared case against your policy; acceptance thresholds for review time, evidence completeness, critical errors and appropriate escalation are agreed before testing begins. The report at the end shows measured benefits, observed failures, remaining dependencies and the work left before production could be considered.

The sector briefs with the sample review package: Banking: periodic KYC reviews · Insurance: motor claims

Contact: contact@eigennetworks.com

Data behind the charts

Frontier list prices, cost per case and per month for each desk
Provider and modelUSD per 1M: in / cached / outClaims desk, EUR per caseClaims desk, EUR a monthKYC desk, EUR per caseKYC desk, EUR a month
OpenAI GPT 5.55.00 / 0.50 / 30.000.1975,9240.2286,838
Claude Opus 55.00 / 0.50 / 25.000.1785,3260.2056,154
Gemini 3.1 Pro2.00 / 0.20 / 12.000.0792,3700.0912,735
Claude Sonnet 52.00 / 0.20 / 10.000.0712,1300.0822,462
Copilot Studiotool call 5 credits, answer 2, at 0.01 USD0.33710,1230.40112,028

Read from the OpenRouter catalogue on 5 October 2026; cached input at one tenth of list as the providers publish; 1.138 USD per EUR. Copilot Studio at 0.01 USD per credit, 5 credits per agent action and 2 per generative answer, pay as you go; Microsoft 365 Copilot seats are not a per case price and are left out.

Workflows: tool calls, runs a day, tokens and frontier cost per case
WorkflowTool callsRuns a dayFrontier tokensEigen N tokensGPT 5.5 EURSonnet 5 EUR
Insurance claims desk
C1 Clean claim within limit, pay540064,35040,0000.1750.063
C2 Invoice missing but obtainable610076,25045,0000.1950.070
C3 Exclusion applies, reject68076,25045,0000.1950.070
C4 Repair above sum insured, partial712089,10050,0000.2160.078
C5 Estimate inconsistent with history, refer1050133,35065,0000.2820.102
C6 Loss before cover incepted, reject65076,25045,0000.1950.070
C7 Peril sub limit, partial (measured)880102,90055,0000.2380.086
C8 Estimate pending, block56064,35040,0000.1750.063
C9 Estimate filed as document, pay76089,10050,0000.2160.078
Finance KYC desk
R Periodic KYC review770089,10050,0000.2160.078
T1 Clean SME onboarding8120102,90055,0000.2380.086
T5 Clean exporter, complete file860102,90055,0000.2380.086
T8 Stale incorporation documents945117,65060,0000.2600.094
T4 Document and register mismatch1040133,35065,0000.2820.102
T3 PEP linked owner1115150,00070,0000.3050.110
T6 PEP ultimate beneficial owner1110150,00070,0000.3050.110
T2 Sanctioned beneficial owner105133,35065,0000.2820.102
T7 Sanctioned corporate owner115150,00070,0000.3050.110
GPU classes: server price, capacity and monthly cost per desk
GPU classCardsServer EUR: purchase / a monthCases a day per server, claims / KYCDesk EUR a month, claims / KYC
24 GB datacenter (L4 class)210,000 / 4491,004 / 830449 / 898
24 GB baseline (RTX 4500 Ada / A10G)210,600 / 4991,674 / 1,383499 / 499
32 GB consumer (RTX 5090)29,800 / 5744,186 / 3,459574 / 574
48 GB workstation (RTX 6000 Ada)114,500 / 5941,339 / 1,107594 / 594
96 GB workstation (RTX PRO 6000)114,500 / 6331,506 / 1,245633 / 633
Rented 24 GB class2rented / 1,1401,674 / 1,3831,140 / 1,140
Rented 48 GB class1rented / 6361,339 / 1,107636 / 636

Eigen Networks UG · contact@eigennetworks.com · Benchmark findings, October 2026 · Numbers reproduce from the assumptions listed on this page.