claude sonnet 5 is out and it's cheap

anthropic dropped Sonnet 5 on July 1. it's now the default model for free and pro users.

the headline isn't raw intelligence. it's price to performance. it's landing close to Opus 4.8 on a bunch of tasks while costing a fraction, and pricing is low through end of august.

on agentic coding benchmarks it jumped from 58% to 63%. on terminal and command line tasks it jumped over 20 points compared to Sonnet 4.6. basically it got noticeably better at just doing stuff on its own instead of waiting for you to babysit every step.

anthropic is betting most people don't need the biggest model, they need something that can run tools, browsers, and terminals without burning the monthly budget in a week. a lot of teams got burned in Q2 when their agent bills spiraled, so this is the direct answer to that.

if you're building anything agentic, coding assistants, automation, browser agents, it's worth testing this week while pricing is still introductory.

fable 5 and mythos 5 are back

quick catch up if you missed it. anthropic released Fable 5 and Mythos 5 in june, then had to pull access for both a couple weeks later because of US export controls. classic "here's your new toy, actually give it back" moment.

the government lifted those controls on june 30 and anthropic restored access on july 1. so that whole saga is resolved for now, though there's still chatter about whether the restored version comes with tighter restrictions like id verification and usage caps compared to what people originally signed up for. worth keeping an eye on if you were using either model.

local AI is having a moment

there's a growing thread across AI communities about people ditching cloud subscriptions and running models locally instead. the "why am i paying $80 a month for something my own GPU can do" energy is real right now.

ollama paired with a frontend like Open WebUI is the combo people keep bringing up. it lets you run open models like Qwen on your own machine with decent accuracy and basically zero ongoing cost.

installers have gotten a lot less painful too, so it's not just a hobbyist thing anymore.

the money and infra side

a few things worth knowing even if you're not deep in the business side of AI.

meta is building out its own cloud compute business to rent AI infrastructure directly against AWS and Azure, basically monetizing the massive amount of compute it's buying this year.

cloudflare set a september deadline requiring AI companies to separate their search crawlers from training crawlers or get blocked by default on publisher sites. that's going to force a lot more licensing deals out into the open instead of happening quietly behind closed doors.

reddit's own AI licensing deals with google and openai are up for renewal this year too, which is part of why the stock has been swinging around lately.

a few tools worth trying this week

claude sonnet 5. if you're already in the claude ecosystem, just try it. it's the default now on free and pro, no extra setup needed, and the terminal and agentic improvements are noticeable if you're doing any coding work.

ollama + open webui. want to try running models locally without a monthly subscription? this combo is the easiest entry point right now. good for privacy sensitive work too since nothing leaves your machine.

perplexity. still the community favorite for research and fact checking when you don't fully trust a chatbot's answer.

aider. solid pick if you want an AI pair programmer that works directly in your terminal against your actual codebase instead of a chat window.

whisperx. if you ever need to transcribe audio or video with speaker labels and timestamps, this is faster and more accurate than plain whisper.

that's the roundup 🎬 the big theme this week is cost and control. cheaper agentic models, cheaper local alternatives, and more scrutiny over who gets to use frontier AI and under what terms.