BYOK
SESSION

A hands-on walkthrough for choosing and wiring your own model.
The cost, quality, and routing decisions an applied-AI engineer has to make before shipping — run live, not slideware.

byok multi-model-routing cost-vs-quality hands-on

section · 01/04

cat why_now.md

// Why Now

Model prices and quality shift every few weeks — Opus 4.6, GPT-5.1, and Gemini 2.5 Pro all moved last quarter alone. Teams that hardcode one vendor either overpay or fall behind. Open Design is local-first · BYOK · Apache-2.0 — bring your own keys, swap providers per task, keep every routing decision in your own hands.

workflow · byok routing

The BYOK workflow, mapped

Four steps from a raw API key to a routed, budget-capped model roster

connect

4 keys

Anthropic, OpenAI, Google, Azure — added once in Settings, never leave your machine.

compare

3 models

Same prompt, same task — latency, cost, and output quality side by side.

route

1 policy

Task-type rules decide which model answers — no manual switching mid-project.

// keep

byok = your cost, your data

Every call rides your own key; nothing routes through a shared subscription.

// avoid

single-vendor lock-in

One outage or price hike shouldn't stall the whole team's roadmap.

trace · pilot run

od model pilot --tasks 42 --models opus-4.6,gpt-5.1,gemini-2.5-pro

↓ Two-week pilot, three teams, one dashboard

# open design · pilot session 07c3
[connect] anthropic + openai + google keys   // added once, stored locally
[route]   "long-context spec review → gemini-2.5-pro"
[route]   "quick UI copy pass → gpt-5.1-mini"
[route]   "complex agent planning → claude-opus-4.6"
[measure] cost/task, latency, pass-rate      per model, per task type
[report]  pilot-summary.md                  PASS 42/42 tasks routed

# 2 weeks · 3 teams · 42 tasks · zero vendor lock-in

benchmark · spend variance

Risk control: -71% cost variance

vs single-vendor default: day-to-day spend swings tamed by routing guardrails (pilot data, 3 teams)

tldr

echo $ROI

3.1x roi

42 tasks, 3 models, 2 weeks — routed spend down 46%, output quality held within 2 points of the best single-model baseline.

+ what moved

• cost down 46% vs single-vendor baseline
• latency down 31% avg — simple tasks route to a faster model
• zero missed SLA across the pilot

+ what's included

• 4-provider BYOK, keys stored locally
• per-task-type routing policy
• budget caps + audit log on every call

install

Roll it out to your team

Three commands to stand up your own BYOK routing pilot.

# 1. clone (apache-2.0)
$ git clone https://github.com/nexu-io/open-design

# 2. add your keys (byok, stored locally)
$ od keys add anthropic openai google

# 3. start a routed pilot
$ od model pilot --tasks ./pilot-tasks.json

local-first byok budget-caps

EOF

exit 0

// thanks

Full pilot data, routing policy templates, and the cost dashboard live at github.com/nexu-io/open-design · discord.gg/mHAjSMV6gz · x.com/nexudotio