Explore the knowledge library
This archive contains only material that is currently maintained. Pages come from scoped practical work or source-bounded research; material that cannot be reviewed reliably is revised, replaced, or removed from the public index.
How to Estimate LLM API Cost Before Shipping: A Token Budget You Can Verify
Turn measured input, cached-input, and output tokens into a per-request and monthly budget, then reconcile the forecast against provider usage and billing data.
Kimi K3.1 Pro Explained: Long Context, Coding Claims, Pricing, and a Safer Test Plan
Moonshot's K3.1 Pro combines a 3T-class model, a one-million-token context window, and a faster variant. This guide separates official claims from what a real engineering evaluation still needs to measure.
OpenAI Plans to End Cursor Model Access: What Developers Should Prepare Before November 12
OpenAI says it intends to wind down its Cursor model contract after SpaceX acquired the coding-tool company. Here is what is confirmed, what remains uncertain, and how teams can reduce model-provider lock-in now.
Tencent Hy4 Preview Explained: What 770B and a 1M Context Window Actually Change
Tencent has released and open-sourced Hy4 preview. Here is what the official release says about its 770B/49B MoE design, one-million-token context, access, pricing, and the evidence developers should still verify.
OpenAI’s Private Safety Processing: What Zero Data Retention Changes for Builders
OpenAI is previewing a safety layer for Zero Data Retention deployments. Here is what the announcement says, what it does not promise, and what teams should verify before relying on it.
Claude Text Watermarking Explained: Detection, Limits, and What It Means for Writers
Anthropic says future Claude models will use invisible statistical watermarks. The useful question is not whether a detector can label every paragraph, but what the signal can and cannot establish.
Build a Repeatable LLM Agent Evaluation Before You Pick a Model
A practical evaluation design for comparing coding and tool-using models on the work your product actually needs.