Verdict (30-second read): If your lakehouse is bleeding cash because your "cold tier" is still calling GPT-4.1 every night, route the summarization job — not the interactive tier — through an OpenAI-compatible aggregator like HolySheep AI using DeepSeek V3.2 at $0.42 / MTok output. You keep the same API contract, swap one URL, and a typical 50 MTok / month cold-tier bill falls from $400 → $21 without a meaningful quality hit on summary-style work.
Provider Comparison — 2026 Metered Output Pricing (public)
| Provider | GPT-4.1 / MTok out |
Claude Sonnet 4.5 / MTok out |
DeepSeek V3.2 / MTok out |
Gemini 2.5 Flash / MTok out |
Typical p50 | Billing & Payment | Best Fit |
|---|---|---|---|---|---|---|---|
| OpenAI (direct) | $8.00 | — | — | — | ~210 ms | USD card only | Interactive NA/EU serving |
| Anthropic (direct) | — | $15.00 | — | — | ~340 ms | USD card only | Long-context premium |
| Google AI (direct) | — | — | — | $2.50 | ~190 ms | USD card only | High-volume cheap sidecar |
| OpenRouter | $8.00 | $15.00 | $0.42 | $2.50 | ~280 ms | USD card only | Model-hopping prototypes |
| HolySheep AI | $8.00 | $15.00 | $0.42 | $2.50 | <50 ms (DPS region) | ¥1=$1 internal rate, WeChat / Alipay / Card | LTAP cold-tier summarization, APAC teams |
Pricing is per each vendor's 2026 public page. The <50 ms p50 figure for HolySheep is published on their DPS-1 endpoint profile; the others are the typical published p50 ranges for cross-region egress we observed in 2025–2026.
What "LTAP" Actually Means Here
Layered Table Access Platform (LTAP) is the pattern I use with clients who run Iceberg-on-S3 or Delta-on-S3 with three explicit zones:
- Hot tier — last 24 hours, queried by BI tools. Stays on a real-time serving model.
- Warm tier — last 30 days. Materialized summaries refreshed on a cron.
- Cold tier — everything else, sitting as Parquet row groups on S3 Standard-IA or Glacier IR. This is where summarization gets expensive if you over-pay for it.
The mistake I see repeatedly: teams point the cold Parquet summarizer at GPT-4.1 because "it's the model we already trust." For a 3-bullet narrative roll-up of a 2 k-token row group, you do not need an $8/MTok model. You need a $0.42/MTok model with strict latency SLAs and predictable per-token cost.
The 2026 Cost Math (Concrete)
Assume your cold