ShortFlix — $500 Credit Budget (the-cost-conscious)

ShortFlix · ledger

ledger # SF-2026-001
billing acct · gcp-startup-credits
FX · USD
period · 2026-05-06 → 2026-06-05

PROJECTED 30-DAY SPEND

$386.10

of $500.00 startup credit · 77.2% used · $113.90 reserve

Gemini API Cloud Run Cloud SQL Vertex AI Search CDN/Logs/Storage

line item	plan	unit cost	qty	subtotal
Gemini 2.0 Flash · curator + searchstreaming	pay-as-you-go	$0.075/1Mtok in	1.6 Mtok	$120.00
Cloud Run · ADK orchestratormin=1	regional ANE3	$0.024/vCPU-hr	744 hr	$72.00
Vertex AI Search · grounding1 corpus	std edition	$2/1k queries	28k	$56.00
Cloud SQL · curator-cachedb-f1-micro	shared-core	$1.20/day	30 d	$36.00
RapidAPI · YT/IG/TT bundles3rd-party	not on GCP	$25/mo × 3	1 mo	$75.00
Cloud CDN + egress	global	$0.08/GB	200 GB	$16.00
Logging / Monitoring	std	incl.	—	$8.00
Buffer · model overruns / re-renders	contingency	—	—	$3.10
TOTAL · 30-day demo window				$386.10

Cost-Cutting Tactics that Don't Hurt Judging

1 · Flash over Pro

Use gemini-2.0-flash for curator + search. Reserve gemini-2.0-pro only for trend-safety reasoning step (5% of calls).

saves ≈ $140 / month

2 · Cache > Recompute

Curator runs nightly on Cloud Scheduler, not per-request. Per-user re-rank is cheap; embedding is expensive.

saves ≈ $90 / month

3 · Cloud Run min=1

Avoid GKE entirely (over-engineered for demo). Single Cloud Run service, scale-to-1, not zero (judge cold-start).

saves ≈ $200 vs GKE

4 · MCP tool call dedup

Hash query → 60s in-memory cache for RapidAPI fanout. Identical search hits return free.

saves ≈ $30 / month

5 · Vertex Search 1-corpus

Single grounding corpus (cross-cultural metadata) instead of 3. Cheaper, still kills hallucination.

saves ≈ $60 / month

6 · Free-tier defaults

Cloud SQL f1-micro · Logging 50GB free · CDN signed-URL caching · all default-on.

stays under credit ceiling

JUDGING TIE-IN · Business Case 30% — "would this survive at scale?"
Answer: at 10k DAU we project $0.038 / DAU / month, well below a $4.99 Pro tier. Multi-agent system is cheaper per useful recommendation than a single-agent baseline because the curator caches 90% of work overnight.