Claude's new Usage & Cost API reports spend by model, key, user
The old usage and cost report endpoints go legacy, so wire up billing dashboards now. Plus: Messages API cap hits 128MB, and a CPU baked into Qwen3.5.
Copy markdown
Query your Claude bill from code
The new Usage & Cost Analytics API (beta header analytics-usage-cost-2026-08-24) adds /v1/organizations/analytics/usage and /v1/organizations/analytics/cost, returning token counts and dollar cost you can filter and group by model, workspace, API key, and the user or service account behind each request. Build spend dashboards and per-key budget alerts instead of scraping the Console.
Messages API now takes 128MB requests
Anthropic raised the maximum request size on the Messages API and Token Counting API from 32MB to 128MB — 4x the headroom to pack big PDFs, long transcripts, or large tool outputs into a single call without chunking. Fewer round-trips and simpler pipelines for document-heavy agents.
A RISC-V CPU compiled into Qwen3.5's weights
"Qwenputer" (front page of Hacker News, 662 points) bakes a 32-bit RISC-V CPU and a filesystem straight into Qwen3.5-4B — 12 compiled layers sit ahead of Qwen's 32 and use Gated DeltaNet state as RAM. It gives the model memory that survives across sessions from an empty context: 83.3 vs 7.8 for base Qwen on a multi-session recall test, ~512 files stored, running on a 24GB GPU. A weekend rabbit hole into memory-in-weights instead of an external vector DB.