AI Product Updates: New Features and Releases

SOMA now integrates with Chutes, compressing agent context before inference so each request bills fewer tokens and costs less.

LangChain's LangSmith Fine-Tuning is now in public beta, and its open-source CLI, smithtune, trains models on Baseten Loops.

The AI agent-driven privacy routing tool for swaps and transfers enters beta with 192 trading pairs and asset shielding.

LangSmith has released Custom Apps, now generally available, letting users generate any interface from agent data with a single prompt.

Announced at Meta Connect, Replit now builds apps for Meta devices when you describe them.

LangSmith now supports Trajectories as a first-class feature for analyzing long-running agent runs.

Replicas has released an iOS app that lets users run and check on cloud agents from their phone.

LangChain launched smithtune in public beta, which post-trains agent policies through a training API.

LangSmith released smithtune in public beta, turning agent traces into supervised fine-tuning data with an end-to-end pipeline from curation to training and deployment.

LangSmith launched Fine-Tuning, letting teams fine-tune models on their LangSmith trace data via Fireworks AI and Baseten.

MalaSwap launched what it calls the first agentic AI-powered token launchpad on Arc, letting an AI agent create and launch a token from a single command starting September 27.

Perplexity introduced Photon, which cuts p99 response time from about 800 ms to about 65 ms, uses about 20% fewer serving machines, and stores 2.5x as much data per document.