# Controlling LLM cost

> Cutting token spend and API bills: model routing, prompt and KV caching, batching, context budgets, moving work from models to code.
> 6 recommendations, newest first. Canonical: https://wpnews.pro/playbooks/llm-cost

- Cap loop context or restart sessions from a handoff file of current priorities, and prune the system prompt, since caching was 78% of the bill and pruning would save about $93 a day.  
  — [Agent Swarm Economics: Why Caching Was 78% of Our Bill](https://wpnews.pro/news/agent-swarm-economics-why-caching-was-78-of-our-bill), source: <https://verial.xyz/posts/agent-swarm-economics>, 2026-11-17
- Trim or summarize conversation history before each request instead of re-sending every prior turn, since a 50-message chat can send roughly 19x the input tokens of the first message.  
  — [Why message 50 of your chatbot costs 19x more than message 1](https://wpnews.pro/news/why-message-50-of-your-chatbot-costs-19x-more-than-message-1), source: <https://dev.to/deployedlabs/why-message-50-of-your-chatbot-costs-19x-more-than-message-1-1no3>, 2026-11-17
- Test lighter models against your own tasks before defaulting to a frontier model, and scope agent requests with check-ins; reasoning mode on Mistral Large 4 doubled tokens for no accuracy gain.  
  — [Engineering Leaders Hire For Judgment As AI Agents Take On More Of The Coding Work](https://wpnews.pro/news/engineering-leaders-hire-for-judgment-as-ai-agents-take-on-more-of-the-coding), source: <https://www.thestateofai.com/news/ai-agents-engineering-hiring-judgment-smaller-models-david-woollard-standard-ai>, 2026-11-17
- Weigh the 3-point index gap against per-task price when routing work: GPT-6.2 Astra scored 62 at about $4 per task versus Claude Fable 5.2's 65 at $26.54.  
  — [OpenAI’s GPT-6.2 Astra trails Claude Fable 5.2 by three points on an independent AI index at a fraction of the cost per task](https://wpnews.pro/news/openais-gpt-6-2-astra-trails-claude-fable-5-2-by-three-points-on-an-independent), source: <https://www.rdworldonline.com/openais-gpt-6-2-astra-trails-claude-fable-5-2-by-three-points-on-an-independent-ai-index-at-a-fraction-of-the-cost-per-task/>, 2026-11-17
- Route reasoning-heavy steps to a frontier model and mechanical execution to a cheaper, faster model; Google's Orchestrator matched Carbon's success rate with 24% less quota and 19% faster completion.  
  — [Google's Carbon and Hydrogen Models: How the Orchestrator Cuts AI Costs](https://wpnews.pro/news/google-s-carbon-and-hydrogen-models-how-the-orchestrator-cuts-ai-costs), source: <https://www.mindstudio.ai/blog/google-gemini-carbon-hydrogen-orchestrator/>, 2026-11-17
- Route merge-quality coding to GPT-6.2 Sol at $2/$10 and reserve Astra at $8/$40 for agent deliverables needing slides, spreadsheets or science terminal runs, since Sol trails Astra by only 0.5 points on FrontierCode 2.0.  
  — [GPT-6.2 Astra vs Sol vs Claude Fable 5.2 vs Gemini 4 Argon (November 2026)](https://wpnews.pro/news/gpt-6-2-astra-vs-sol-vs-claude-fable-5-2-vs-gemini-4-argon-november-2026), source: <https://explainx.ai/blog/gpt-6-2-astra-sol-vs-claude-fable-gemini-comparison-november-2026?utm_source=rss&utm_medium=feed&utm_campaign=blog>, 2026-11-17
