ATOM AIPI Weekly, Week 12 (March 30, 2026)

[Steriani]

Week 12 numbers are out.

2,614 SKUs across 1,234 models and 47 vendors. All 14 AIPI indexes came in flat this week, 0.0% WoW across every modality and channel category.

Numbers worth knowing:

Text input benchmark (AIPI TXT GLB): $0.001190 per 1K tokens

Text output benchmark: $0.005540 per 1K tokens

Output premium over input: 3.85x

Open source vs proprietary median: OSS is 70.7% cheaper

Caching savings where available: 69.7%

Models offering cached pricing: 20.4%

Context cost curve (128K+ vs 32K): 3.0x

Three straight weeks of flat pricing. The competition right now is on model breadth, not price.

We publish a free MCP server with four tools that work out of the box in Claude, Cursor, and Windsurf with no API key. MCP PRO at $49/mo unlocks search_models, get_model_detail, compare_prices, and get_vendor_catalog.

[Steriani]

Full data and weekly indexes published every Monday. MCP server available on npm and Smithery for anyone building with Claude, Cursor, or Windsurf.

[Steriani]

New data point from this week’s index. Buying inference direct from the model developer costs 7x more per input token than buying the same model through a third-party inference platform. The gap widens to 5.2x on output tokens. Most teams pay this premium by default simply because they never compare across channels.