Decision Memo

Cost router that picks the cheapest capable AI model

Record the team verdict, rationale, and reviewer leans locally, then print or share a source-anchored memo.

Markdown export

Agent and email version.

# Decision Memo: Cost router that picks the cheapest capable AI model

Full report: https://ideanavigatorai.com/ideas/cheapest-ai-model-router/
Recorded: Not recorded

## Decision
- Team verdict: Park
- Validation verdict: Research (59/100)
- Confidence: 55%
- Recommendation: Keep this parked until the team has evidence for the next validation step: Take three companies' last month of API logs, replay them through the router in shadow mode, and present the audited savings-versus-quality delta as the sales artifact.

## Team rationale
No team rationale recorded yet.

## Reviewers
- No named reviewers recorded.

## Source anchors
- Buyer: Engineering lead at a company whose LLM API bill is growing faster than usage
- Market: AI infrastructure / LLM ops
- Problem: Teams route every request to a flagship model by default, paying 10-30x more than necessary for tasks a small model handles identically, because per-task capability testing is tedious and model prices change monthly.
- Thesis: Cost router that picks the cheapest capable AI model should be tested as a narrow first-win workflow for Engineering lead at a company whose LLM API bill is growing faster than usage.
- Source: https://openrouter.ai/
- Source: https://docs.claude.com/

## Validation rubric
Rubric version: INAV-VALIDATION-2026-06-04

### Demand signal - 5.5/10 (24% weight)
Demand looks thin because the report has 2 source-backed signal(s), an editorial confidence of 55/100, and a defined buyer in AI infrastructure / LLM ops.

- Price-per-token differs by more than an order of magnitude between model tiers that score identically on many production task types.
- Target buyer: Engineering lead at a company whose LLM API bill is growing faster than usage

### Problem severity - 6.3/10 (22% weight)
Problem severity is thin when the buyer pain, customer value, and dream-outcome scores are combined.

- Teams route every request to a flagship model by default, paying 10-30x more than necessary for tasks a small model handles identically, because per-task capability testing is tedious and model prices change monthly.
- Price-per-token differs by more than an order of magnitude between model tiers that score identically on many production task types.

### Willingness to pay - 5.5/10 (20% weight)
Willingness to pay is weak; the model has a monetization hypothesis, but it must still be proven through paid pilots or explicit pricing objections.

- Percentage-of-savings pricing or flat platform fee per routed volume tier.
- Take three companies' last month of API logs, replay them through the router in shadow mode, and present the audited savings-versus-quality delta as the sales artifact.

### Competitive saturation - 6/10 (18% weight)
No source-backed direct match is recorded yet, so saturation risk is treated as unknown rather than proof of novelty.

- Existing-product check has no named direct match.
- Competitive score rewards a narrow wedge, not absence of research.

### Feasibility - 6.2/10 (16% weight)
Feasibility is thin for a moderate build if the MVP is limited to the first measurable workflow.

- Take three companies' last month of API logs, replay them through the router in shadow mode, and present the audited savings-versus-quality delta as the sales artifact.
- OpenRouter and provider-native routing features already occupy adjacent ground.

## Market gap
Underserved segments:
- Engineering lead at a company whose LLM API bill is growing faster than usage who still run the workflow in spreadsheets, generic docs, email, or chat threads.
- Small teams in AI infrastructure / LLM ops that feel the pain weekly but are too narrow for broad incumbents.
- New adopters who need guided proof before committing to a larger platform.

Feature gaps:
- A narrow workflow that reaches value without configuration-heavy onboarding.
- A buyer-facing proof artifact that shows time saved, risk reduced, or communication improved.
- A handoff path from manual concierge service to repeatable software.

Differentiation levers:
- Use specificity as the wedge: one buyer, one workflow, one measurable result.
- Show proof earlier than broad competitors with before-and-after examples and small pilot data.
- Keep implementation lighter than incumbent suites or generic AI assistants.

## Roast and risks
Promising enough to test, not strong enough to build broadly.

Blind spots:
- OpenRouter and provider-native routing features already occupy adjacent ground.
- A broad AI assistant can flatten differentiation unless the wedge is painfully specific.
- The first release can become a generic dashboard if the job is not named tightly.

Hard questions:
- Who wakes up already trying to solve this?
- What do they stop paying for or stop doing when this works?
- What proof would make a skeptical buyer trust it in one screen?
- What is the smallest paid version of this idea?

## Kill criteria
- Fewer than five qualified buyers agree to discuss the workflow after targeted outreach.
- No buyer can name a current cost in time, money, risk, or reputation.
- The first demo does not produce a clear next step, paid pilot, or specific objection.

## Offer ladder
- **Lead magnet (Free)**: Cost Router That Picks The Cheapest Capable Ai Model checklist Goal: Capture qualified leads and learn the buyer's exact language. Value: Helps Engineering lead at a company whose LLM API bill is growing faster than usage audit the painful workflow before buying software.
- **Frontend offer ($19-$99)**: Concierge review or paid template Goal: Validate urgency, workflow fit, and willingness to pay. Value: Delivers the first useful output manually before automation is trusted.
- **Core offer ($49-$499/month)**: Cost router that picks the cheapest capable AI model focused SaaS Goal: Create the recurring revenue product after the narrow wedge survives tests. Value: Turns the recurring manual workflow into a repeatable product loop.
- **Continuity ($99-$1,000/year add-on)**: Monitoring, benchmarks, and monthly reporting Goal: Increase retention and make the product part of a routine. Value: Keeps the buyer engaged with ongoing proof, saved time, or reduced risk.
- **Backend offer (Custom)**: Done-with-you setup, agency, or team rollout Goal: Capture higher-value accounts once the productized wedge is proven. Value: Adds implementation help, integrations, and workflow migration.

Print-ready memo

Decision Memo: Cost router that picks the cheapest capable AI model

Team verdict: Park
Validation verdict: Research / 59/100
Confidence: 55%
Recorded: Not recorded

Recommendation

Keep this parked until the team has evidence for the next validation step: Take three companies' last month of API logs, replay them through the router in shadow mode, and present the audited savings-versus-quality delta as the sales artifact.

Team rationale

No team rationale recorded yet.

Reviewers

No named reviewers recorded.

Source anchors

Buyer: Engineering lead at a company whose LLM API bill is growing faster than usage
Market: AI infrastructure / LLM ops
Problem: Teams route every request to a flagship model by default, paying 10-30x more than necessary for tasks a small model handles identically, because per-task capability testing is tedious and model prices change monthly.
Thesis: Cost router that picks the cheapest capable AI model should be tested as a narrow first-win workflow for Engineering lead at a company whose LLM API bill is growing faster than usage.

Validation rubric

Demand signal

24% weight

5.5/10

Demand looks thin because the report has 2 source-backed signal(s), an editorial confidence of 55/100, and a defined buyer in AI infrastructure / LLM ops.

Problem severity

22% weight

6.3/10

Problem severity is thin when the buyer pain, customer value, and dream-outcome scores are combined.

Willingness to pay

20% weight

5.5/10

Willingness to pay is weak; the model has a monetization hypothesis, but it must still be proven through paid pilots or explicit pricing objections.

Competitive saturation

18% weight

6/10

No source-backed direct match is recorded yet, so saturation risk is treated as unknown rather than proof of novelty.

Feasibility

16% weight

6.2/10

Feasibility is thin for a moderate build if the MVP is limited to the first measurable workflow.

Market gap

Underserved segments

Engineering lead at a company whose LLM API bill is growing faster than usage who still run the workflow in spreadsheets, generic docs, email, or chat threads.
Small teams in AI infrastructure / LLM ops that feel the pain weekly but are too narrow for broad incumbents.
New adopters who need guided proof before committing to a larger platform.

Feature gaps

A narrow workflow that reaches value without configuration-heavy onboarding.
A buyer-facing proof artifact that shows time saved, risk reduced, or communication improved.
A handoff path from manual concierge service to repeatable software.

Differentiation levers

Use specificity as the wedge: one buyer, one workflow, one measurable result.
Show proof earlier than broad competitors with before-and-after examples and small pilot data.
Keep implementation lighter than incumbent suites or generic AI assistants.

Roast and risks

Promising enough to test, not strong enough to build broadly.

Blind spots

OpenRouter and provider-native routing features already occupy adjacent ground.
A broad AI assistant can flatten differentiation unless the wedge is painfully specific.
The first release can become a generic dashboard if the job is not named tightly.

Hard questions

Who wakes up already trying to solve this?
What do they stop paying for or stop doing when this works?
What proof would make a skeptical buyer trust it in one screen?
What is the smallest paid version of this idea?

Kill criteria

Fewer than five qualified buyers agree to discuss the workflow after targeted outreach.
No buyer can name a current cost in time, money, risk, or reputation.
The first demo does not produce a clear next step, paid pilot, or specific objection.

Offer ladder

Lead magnet