Anthropic's flagship model gets a significant capability jump. Claude 4 Opus ships with a 200k-token context window, native tool use improvements, and — the headline feature — an "extended thinking" mode. When you enable it, the model reasons out loud through sub-steps before committing to an answer, dramatically improving accuracy on multi-step math, code review, and research synthesis. Pricing: $15/M input tokens, $75/M output tokens in standard mode. Extended thinking charges for the reasoning tokens separately. Available now in the API and for Claude.ai Pro subscribers.
Perplexity's web search AI finally gets a purpose-built mobile app — not just a wrapper. The Assistant app adds voice input with real-time transcription, a persistent conversation history that syncs across devices, and a new "Deep Research" mode that runs multi-step web searches in the background and returns a structured report. The free tier allows 5 deep searches per day; Pro ($20/mo) is unlimited. Early reviews on the App Store suggest the voice experience is snappier than asking Siri a follow-up question.
Vercel ships a major version of its AI SDK with first-class support for multi-provider routing, structured outputs with Zod validation, and a new streaming UI primitives API. The standout: a built-in provider router that can fall back from Claude to GPT-4o to Gemini automatically based on latency or rate limits. Particularly useful for production apps that can't afford downtime during a single-provider outage. Ships with adapters for Next.js, SvelteKit, and Nuxt. Available on npm now.
LangChain's agent orchestration runtime exits preview. LangGraph Cloud now includes a visual debugger for tracing agent execution paths, one-click horizontal scaling, and native checkpointing so long-running agents can resume after interruption. The managed tier starts at $49/mo; the open-source runtime is still MIT-licensed. The most requested feature — a UI for non-engineers to trigger and monitor agents without writing code — is shipping in beta alongside the GA launch.
Joins your Google Meet or Zoom call silently, transcribes in real time, and pushes live cue cards to your screen when it detects you might be dominating the conversation or drifting off-topic. Post-meeting, it generates a structured summary with action items and sentiment analysis. Works without installing anything on attendees' machines — only the host needs the extension.
Upload a photo of any room and get 10 redesigned variants in under 30 seconds — different furniture styles, paint colors, and lighting setups. The $12 tier adds an "ask anything" chat that lets you describe what you want changed ("remove the rug, add a reading nook by the window") and regenerates accordingly. Impressive quality for the price; the free tier gives you 3 renders/day.
Build multi-step agents with a drag-and-drop canvas. Connect to Gmail, Notion, Airtable, Slack, and 40+ other tools. Each "node" in the flow is a prompt you write in plain English — FlowAgent handles the API calls, error handling, and retry logic. The free tier allows 1 active agent; paid plans ($19+/mo) unlock parallel runs and custom integrations. Competes with Make and Zapier AI, but targets developers who want more control without writing code.
Connect your database (Postgres, MySQL, BigQuery, Snowflake) and ask questions in plain English. SQLMind writes the query, explains what it does, and lets you run it in one click. The standout feature is "query history context" — it learns your schema and typical query patterns over time to give more relevant suggestions. SOC 2 Type II certified; queries never leave your browser.
Upload PDFs, Word docs, or paste a URL and get an AI that can answer specific questions with citations to the exact page and paragraph. Handles documents up to 500 pages on the free plan. The Pro tier adds team sharing, API access, and a "compare documents" mode that surfaces key differences across multiple files. Good for legal, finance, and research teams.
Clone your voice from a 30-second sample and use it in text-to-speech, video narration, or podcast intros. Notably, VoiceClone requires identity verification before you can clone anyone's voice, and adds an inaudible watermark to all generated audio. Output quality is competitive with ElevenLabs at a lower price point. Supports 28 languages for the cloned voice.
A surprisingly nuanced thread — the OP describes replacing a 2,000-line Spark job with a Claude 4 Sonnet prompt that processes CSVs, cleans data, and outputs structured JSON. The top comments range from "this is a terrible idea in production" to "we did the same thing and it's fine." The key insight buried in the thread: for jobs that run <1M rows and where occasional errors are acceptable, LLM-based pipelines are genuinely competitive on cost and development speed.