7/ Stop using Opus for Haiku tasks Opus is roughly 5x more expensive per task than Sonnet. Sonnet is more expensive than Haiku. Quick reply, formatting fix, brainstorm? Haiku.
Writing, analysis, code, real drafts? Sonnet. Deep research, complex reasoning, long document review?
LLMS
-
Optimize AI Model Selection by Task Complexity and Cost
By
–
-
Optimize Claude: Disable Unused Features to Save Tokens
By
–
6/ Turn off features you are not using Web search, research mode, and connectors add tokens to every response even when you do not need them. If you are writing your own content or asking a knowledge question Claude already knows, turn them off. The icons are right next to the
-
Stack Questions for Better AI Responses
By
–
3/ Stack your questions into one message Three separate prompts cost three full reads of your conversation history. One prompt with three questions costs one read. The answers are usually better too because Claude sees the full goal at once. Example: "I need three things from
-
LLM and RAG Applications Guide: LangChain Python Tutorial
By
–
"Generative AI and RAG for Beginners: A Practical Step-by-Step Guide to Building LLM and RAG Applications with LangChain and Python" Get your copy at https://
amzn.to/3MZZ9R5 Independently published: December 2025
Print length: 255 pages -
Optimize Claude conversations by starting fresh every 20 messages
By
–
2/ Start a new chat every 20 messages By message 30, a single question can cost 50,000 tokens because Claude is reprocessing everything that came before it. When the chat gets long, ask Claude to summarize and continue fresh.
Prompt: "Summarize the key points of this -
Edit the Prompt Rather Than Restart the Conversation
By
–
1/ Edit the prompt instead of asking again Every follow-up message makes Claude re-read the entire conversation. The longer the chat, the more tokens each new message burns. When Claude misses what you wanted, click the edit icon on your original prompt, fix it, and regenerate.
-
10 Prompts to Supercharge Claude — Prompt Engineering Tips
By
–
Arrêtez de dire à Claude : "construis ça" Arrêtez de dire à Claude : "écris du bon code" Arrêtez de dire à Claude : "corrige ce bug" Vous utilisez une IA niveau staff engineer comme un stagiaire junior. Voici 10 PROMPTS pour rendre Claude SURPUISSANT [ Ajoutez en signet
-
IBM Granite 4.1 Language and Speech Models Now Available
By
–
Try Granite 4.1 here Language: https://
replicate.com/ibm-granite/gr
anite-4.1-8b
… Speech: https://
replicate.com/ibm-granite/gr
anite-speech-4.1-2b
… -
Mistral Launches Medium 3.5: 128B Dense Model with Vision
By
–
Congrats to the @MistralAI team on launching Mistral Medium 3.5! This new single 128B dense text-vision model merges instruction-following, reasoning, and coding. Plus, it scored 77.6% on SWE-Bench Verified. The best models run on NVIDIA. Try it out today on
-
Grok Speech: Advanced Text-to-Speech and Speech-to-Text AI
By
–
Grok Speech by @xai
— Replicate (@replicate) 29 avril 2026
Text-to-Speech: 5 voices, 20 languages, expressive tags like [laugh] and <whisper>
Speech-to-Text: 25 languages, word-level timestamps, speaker diarization
Try it here 👇 pic.twitter.com/T646xffeGqGrok Speech by @xai Text-to-Speech: 5 voices, 20 languages, expressive tags like [laugh] and Speech-to-Text: 25 languages, word-level timestamps, speaker diarization Try it here