LightNews — News, ranked by scholars

500x Engineer

@500xengineer.bsky.social

I read a week of my own Claude Code logs: 6.5 billion tokens, and 95% were cache re-reads. The longest 10% of threads used 74% of the total.

New long-form on where it went and the fix I'd do first: 500xengineer.com/blog/where-claude-code-tokens-go

#ClaudeCode #AIAgents #PromptCaching

Where Do Claude Code Tokens Actually Go? I Measured 6.5 Billion | 500x Engineer

I read a week of my own Claude Code logs: 6.5 billion tokens, 95% of them cache re-reads. Here's where the tokens went, what list prices say about which ones weigh most, and the fixes I'd do first.

www.500xengineer.com

November 8, 2026 at 4:18 AM

Zarar

@zarar.dev

Where do your tokens actually go? I cover prompt caching, startup costs, /clear vs /compact differences and when to use what, and look at how to optimize memory.

#llm #promptcaching #codingagent

www.youtube.com/watch?v=SMm0...

Where Your Claude Code Tokens Actually Go

YouTube video by Zarar Siddiqi

www.youtube.com

November 2, 2026 at 1:23 PM

MonkeysCode

@monkeyscode.bsky.social

Prefill dropped ~170s to 1.6s on 28k tokens: broken KV caches burn budgets. MonkeysCode's built-in caching fixes this. How do you handle context bloat?

https://monkeyscode.com/features?utm_source=bluesky&utm_medium=company&utm_campaign=story-fbf88dda

#promptcaching #rustlang

November 2, 2026 at 1:02 AM

Chris Rodriguez

@chriserod.bsky.social

Pick the longest contiguous leading stable prefix, verify it meets the model's min cacheable tokens, and use explicit-only if stable content appears after volatile turns. #promptcaching

https://chriseugenerodriguez.com/blog/prompt-prefix-caching-decision

AI-generated illustration of a reusable LLM prompt prefix boundary between stable instructions and a changing request suffix.

November 1, 2026 at 11:50 PM

Jacob GPTProto

@gptproto-jacob.bsky.social

Same Claude Fable 5.1 task: $7.44 with cache, $27 without.

Across 26K Agent Arena sessions, billing
cached tokens at full input price makes a
top-10 agent task cost 3.6x.

A cache invalidation alone adds 5%.

The price list is not the bill. #PromptCaching

October 30, 2026 at 6:21 PM

Joe

@joeoftex.bsky.social

Reading a novel with LLM, making a spoiler free wiki, chapter by chapter! See how I did it using super cheap using Luna and zero thinking! authrot.app/blog/how-aut...
#LLM #AIEngineering #PromptCaching #WritingCommunity

Character sheet in wiki viewing from early chapter

October 27, 2026 at 5:19 PM

MonkeysCode

@monkeyscode.bsky.social

Why let agents re-read the repo? Pi 1.0's Codemode cuts token bloat ~40%. We combine prompt caching & Auto routing to fix this. Try Auto mode.

https://monkeyscode.com/features?utm_source=bluesky&utm_medium=company&utm_campaign=story-1bb612ba

#promptcaching #AIagents

October 21, 2026 at 8:01 PM

Expert AI

@expertai.bsky.social

Оптимизируйте расходы на OpenAI API с помощью Prompt Caching! Сокращайте затраты до 10 раз, кэшируя статичные инструкции и документацию. Главное — соблюдать порог в 1024 токена для максимальной выгоды. #OpenAI #PromptCaching #LLM #Optimization #Python https://ai.dappcrypto.org/r/hG

October 21, 2026 at 4:03 PM

Qiita Trending Bot

@qiita-trending.bot.chrs.to

スキーマを変えるとプロンプトキャッシュが共有されなかったので、Bedrock の構造化出力をやめた話（2026年版）

#AWS #bedrock #Claude #PromptCaching #StructuredOutputs

スキーマを変えるとプロンプトキャッシュが共有されなかったので、Bedrock の構造化出力をやめた話（2026年版） - Qiita

はじめに GMOコネクトの永田です。同じ長い資料を渡して、出力する JSON の形だけを変えながら Claude を何回か呼ぶ、というようなアプリを開発していました。プロンプトの前半（資料の部分）はどの呼び出しでも同じなので、プロンプトキャッシュが効いて、費用も待ち時間...

qiita.com

October 18, 2026 at 1:20 AM

dafu09.bsky.social

@dafu09.bsky.social

还有个隐形收益：稳定的前缀 = 更高的 prompt cache 命中率。

system prompt、工具定义放前面且别动，变化的内容追加在后面。每轮在开头塞个时间戳，缓存就全废了。

省 token 最好的办法，是别让它重复读同样的东西。

#PromptCaching #AIAgent

October 16, 2026 at 5:30 PM

MonkeysCode

@monkeyscode.bsky.social

Anthropic streams timed out during long reasoning phases. We shipped keep-alive pings to fix this, and separated GPT-6 system blocks for cache breakpoints. Tracking cache…

https://monkeyscode.com/features?utm_source=bluesky&utm_medium=company&utm_campaign=story-61869d63

#PromptCaching #DevTools

October 8, 2026 at 3:01 PM

NILE1

@nile11.bsky.social

Anthropic Cuts Claude Haiku 5.5 API Prices by 90%
..................
https://nile1.com/anthropic-cuts-claude-haiku-5-5-api-prices-by-90/
..................
#adjustableEffortSetting #anthropic #apiPricing #claudeHaiku5.5 #darioAmodei #gpt-6Luna #openai #opus5.5 #osworld2.1 #promptCaching #smallLangua

Captura-de-pantalla-2026-10-07-163823.png@webp.webp

October 7, 2026 at 8:51 PM

Marc Mayol

@srmarcmayol.bsky.social

«Cacheamos las llamadas al #LLM» puede ser cuatro cosas. KV cache, prefix caching y #PromptCaching reutilizan cálculo y exigen los mismos tokens. La semantic cache reutiliza respuestas y se fía del significado. Confundirlas sale caro.

Lo estable delante, lo variable detrás.

October 4, 2026 at 10:21 AM

AI Daily Post

@aidailypost.com

OpenAI just dropped GPT‑6, slicing prompt processing time in half and bumping factuality. Plus new Sol, Luna & Astra flavors, smarter caching, and tighter API pricing. Curious how this reshapes inference efficiency? Dive in! #GPT6 #PromptCaching #InferenceEfficiency

🔗

September 25, 2026 at 5:55 AM

donwebmedia.bsky.social

@donwebmedia.bsky.social

Cómo reducir costos de API de LLM en 2026

Descubrí estrategias efectivas para reducir costos api llm en tus proyectos de inteligencia artificial sin perder rendimiento. Descubrilo ahora.

#promptcaching #apillm #gpt56 #costosia #scraping

Cómo reducir costos de API de LLM en 2026

Prompt caching, Batch API, cascadas de modelos y control de scraping bloqueado: los datos concretos de un test publicado por Decodo en septiembre de 2026.

blog.donweb.com

September 25, 2026 at 9:34 AM

Alvin Ashcraft

@alvinashcraft.com

Better prompt caching for GPT‑6 | OpenAI blog

buff.ly/SLApjxt

#ai #openai #gpt6 #models #prompting #promptcaching #tokens #context

Better prompt caching for GPT-6

Learn how GPT-6 improves prompt caching with higher cache hit rates, new diagnostics, explicit breakpoints, and controls that reduce latency and costs.

buff.ly

September 23, 2026 at 7:01 PM

Darryl Ruggles

@darryl-ruggles.cloud

https://lckhd.eu/7tiS3v

#Bedrock #PromptCaching #GenerativeAI #FinOps #MultiTenancy

Almost everyone is watching their AI spend more closely these days. As the price of LLM tokens moves closer to their real cost, teams have to start to optimize spend as much as they can. Prompt caching is one of

Optimizing cost and latency with Amazon Bedrock prompt caching | Amazon Web Services

Prompt caching in Amazon Bedrock can cut input token costs by up to 90% when you repeatedly send the same context to foundation models. This post walks through six practical prompt caching scenarios using the Converse API: message content, system prompt, tool definition, mixed TTL, tenant isolation, and LangChain integration.

lckhd.eu

September 15, 2026 at 6:06 PM

Aleksei Aleinikov

@alekseialeinikov.com

Claude Fable und Mythos kosten doppelt so viel wie Opus. Für Agenten können sie günstiger sein

Die Fable- und Mythos-Linie hat die günstigsten Cache-Reads und die teuersten Cache-Writes im aktiven Lineup. Diese eine Asymmetrie dreht um, welches Modell…

#claudefable #claudemythos #promptcaching

Claude Fable und Mythos kosten doppelt so viel wie Opus. Für Agenten können sie günstiger sein

Die Fable- und Mythos-Linie hat die günstigsten Cache-Reads und die teuersten Cache-Writes im aktiven Lineup. Diese eine Asymmetrie dreht um, welches Modell für Agenten mit langem Kontext günstiger ist — und macht Ihre Cache-Trefferquote viermal wertvoller als anderswo.

www.alekseialeinikov.com

September 4, 2026 at 5:07 AM

Aleksei Aleinikov

@alekseialeinikov.com

Claude Fable and Mythos Cost Twice as Much as Opus. For Agents, They Can Cost Less

The Fable and Mythos line has the cheapest cache reads and the most expensive cache writes in the active lineup. That single asymmetry inverts which model is cheaper for…

#claudefable #claudemythos #promptcaching

Claude Fable and Mythos Cost Twice as Much as Opus. For Agents, They Can Cost Less

The Fable and Mythos line has the cheapest cache reads and the most expensive cache writes in the active lineup. That single asymmetry inverts which model is cheaper for long-context agents, and makes your cache hit rate worth four times more than it is anywhere else.

www.alekseialeinikov.com

September 4, 2026 at 5:07 AM