Inflection-2.5 at 94% of GPT4, and Pi at 6m MAU

**Mustafa Suleyman** announced **Inflection 2.5**, which achieves *more than 94% the average performance of GPT-4 despite using only 40% the training FLOPs*. **Pi**'s user base is growing about 10% weekly, with new features like realtime web search. The community noted similarities between Inflection 2.5 and **Claude 3 Sonnet**. **Claude 3 Opus** outperformed **GPT-4** in a 1.5:1 vote and is now the default for **Perplexity Pro** users. **Anthropic** added experimental tool calling support for Claude 3 via **LangChain**. **LlamaIndex** released LlamaParse JSON Mode for structured PDF parsing and added video retrieval via VideoDB, enabling retrieval-augmented generation (RAG) pipelines. A paper proposed knowledge-augmented planning for LLM agents. New benchmarks like TinyBenchmarks and the **Yi-9B** model release show strong code and math performance, surpassing **Mistral**.