News

• Context window is not max output — and you cannot infer one from the other 2026-07-07

  Two Groq models share the same 131,072-token context window, yet one lets you generate all 131,072 tokens of output and the other caps at 65,536. Context window and max output are independent facts — code that derives one from the other silently truncates responses or throws.

• The same DeepSeek V4 Pro costs 4x more on Together and Fireworks than on DeepSeek's own API 2026-07-07

  DeepSeek V4 Pro is an open-weight model you can call three ways. On DeepSeek's own API it is $0.435 per 1M input tokens; on Together and Fireworks the same model is $1.74 — exactly 4x more. A per-model price is a category error for open-weight models, and at least one widely used cost tracker had the hosted entries at the native price.

• AWS quietly doubled Bedrock prices for legacy Claude models — and cost trackers haven't caught up 2026-07-04

  Amazon Bedrock's 'Public Extended Access' pricing, effective December 1, 2025, doubled per-token prices for Claude 3.5 Sonnet v2. Seven months later, at least one widely used open-source cost tracker still carried the old prices — we verified against the primary source and submitted the fix.

• A facts wire for the model economy: why we exist 2026-07-04

  Software is making purchasing decisions about AI models using unverified data. We built a source-verified feed — every number quoted from a primary source, independently reviewed before publishing.