Tag archive
November 2025, during an architecture audit of a multi-model inference layer for a document search...
A numbers-first 2026 guide to cutting LLM API bills: how prompt caching, batch APIs, and model routing stack across Anthropic, OpenAI, and Gemini.
CLAUDE_CODE_SUBAGENT_MODEL beats every YAML pin. Learn Fable 5 plan requirements, ZDR constraints, and how to declare subagent models withou
Changing context size can reload a local model before every request. A measured RTX 5090 sweep shows why context belongs in the routing key.
How to Cut Inference Costs in Agentic Coding Pipelines with Open-Weight Model Routing and...
Running a 70B parameter model to summarize a 200-word email is wasteful. Running a 3B model to review...
MAX_THINKING_TOKENS, prompt caching, version pinning, Hooks guards, and model routing — seven changes that make high-intensity Claude Code usage financially pre
The full playbook for cheap, sharp agent work. Get observability in first (statusline, burndown, OTEL), compress what reaches the model (rtk, headroom), route who does the work (premium delegates, cheap executes, advisor escalation, model-per-task), then work the long tail: caveman, the right browser validator, scoped scans, fewer MCPs, handover, and a memory system. The capstone of the Token Saving series.
OpenRouter raises $113M Series B led by CapitalG (Alphabet), with NVIDIA, ServiceNow, MongoDB, Snowflake, and Databricks — as the platform now processes
Bernstein's epsilon-greedy bandit router learns which model fits each task type. In our own runs, the bandit cut spend roughly in half. Measure yours with bernstein cost.
SaaStr data shows enterprise AI share shifting hard toward Claude. The lesson isn't pick Claude. It's stop hard-coding one vendor.
One web fetch compounded across 22 agent turns and ate 28% of the entire PR's cost. We added Haiku as a summarization filter to stop the bleed.