Sample savings report

Your AI bill has $4,812/month in testable savings.

SpendLens AI shows which workload to test first, ranks three policy-approved models, and measures cost, speed and reliability before you change production.

First Compare Mode test

Compare three models for product recommendations.

Starting candidates: Gemini 2.5 Flash, DeepSeek V4 Flash and Mistral Small 4. Potential savings: $2,140/month before validation.

Projected monthly spend

$12,654

Based on 18,291 calls yesterday

Potential savings

$4,812/mo

Directional estimate before testing

Ready to evaluate

$3,284/mo

Top candidates available for staging

Payback signal

1 day

Pro plan covered by first accepted switch

Opportunity queue

Workloads worth testing

recommend-products

Strong evidence

shopping/catalog.py::recommend_products

Current: gpt-4o
Starting candidate: Gemini Flash
Spend: $5,210
Savings: $2,140/mo

Compact product-ranking prompts and stable catalog fields make this a strong replay candidate.

summarize-product-reviews

Strong evidence

shopping/reviews.py::summarize_reviews

Current: claude-sonnet-4
Starting candidate: Qwen Turbo
Spend: $3,420
Savings: $1,144/mo

Summaries are under 500 tokens and quality risk is low after replay.

extract-shopping-filters

Moderate evidence

shopping/search.py::extract_filters

Current: gpt-4o
Starting candidate: DeepSeek Chat
Spend: $1,980
Savings: $792/mo

Structured filter extraction looks safe, but needs validation on ambiguous shopping queries.

compare-products

Early evidence

shopping/compare.py::compare_products

Current: claude-sonnet-4
Starting candidate: Llama 3.1 70B
Spend: $2,044
Savings: $736/mo

Open-ended answers need human review before any rollout.

Endpoint	Current	Starting candidate	Spend	Savings	Evidence strength
recommend-products shopping/catalog.py::recommend_products	gpt-4o	Gemini Flash	$5,210	$2,140/mo	Strong evidence Compact product-ranking prompts and stable catalog fields make this a strong replay candidate.
summarize-product-reviews shopping/reviews.py::summarize_reviews	claude-sonnet-4	Qwen Turbo	$3,420	$1,144/mo	Strong evidence Summaries are under 500 tokens and quality risk is low after replay.
extract-shopping-filters shopping/search.py::extract_filters	gpt-4o	DeepSeek Chat	$1,980	$792/mo	Moderate evidence Structured filter extraction looks safe, but needs validation on ambiguous shopping queries.
compare-products shopping/compare.py::compare_products	claude-sonnet-4	Llama 3.1 70B	$2,044	$736/mo	Early evidence Open-ended answers need human review before any rollout.

Savings graph

Monthly savings by workload

Product recommendations$2,140

Review summaries$1,144

Shopping filters$792

Product comparisons$736

$4,812/month total

The first two strong-evidence opportunities cover most of the upside without asking the team to touch risky customer-facing QA yet.

Prompt Waste Signals

Token leaks that do not require a model switch.

These are advisory, rule-based signals from existing metadata. SpendLens does not rewrite prompts automatically or send full user prompts to external services.

Large prompt/context detected

High

Workload: rag-answer

Avg input

18,400

Avg output

640

Review whether all retrieved chunks are needed. Try reducing RAG chunks, filtering context more aggressively, or using prompt caching if available.

Long output detected

Medium

Workload: write-product-description

Avg input

4,900

Avg output

2,300

Add concise-output instructions or max token limits where appropriate.

Large repeated prompt template

Medium

Workload: recommend-products

Avg input

3,200

Avg output

Shorten static instructions or move reusable context outside the repeated prompt.

Evidence strength

Know what to trust first.

High

Move to replay testing now.

Medium

Test carefully on edge cases.

Low

Keep as manual review only.

Suggested next actions

A savings report your engineers can actually ship.

01Replay 200 representative examples for the strongest model shortlist.
02Reduce RAG chunks on rag-answer and compare answer quality against current output.
03Add concise-output instructions to write-product-description and cap max output tokens.
04Pilot the validated model choice on 10% of traffic.
05Watch cost, latency, disagreement, and prompt-token averages for 24 hours.