The headline API price hasn't changed: still $10 per million input tokens, $50 per million output. But in early September, Anthropic released Fable 5.1 — and buried in the pricing is a real change: cache reads dropped from $1.00 to $0.25 per million tokens, a 75% cut. Prompt caching is what happens when you pass the same long system prompt or context on every call — instead of charging you full price each time, Claude reads from the cached version at a fraction of the cost.

For a typical API user, Anthropic says the saving works out to roughly 25% off the bill. For agents and coding workflows — where a long context gets reused on every step — it hits around 45%.

Why you care: if you have Claude running anything automatically — document pipelines, scheduled agents, workflows with a long system prompt — this is money back on the same work, no code change needed once caching is on.

OpenAI released GPT-6 Astra on September 3, and it has been rolling out to ChatGPT Plus, Pro, Business, and Enterprise users since. The headline capability: Astra can use your computer like a person — browse the web, write and run code, fill in forms, build spreadsheets, and create presentations.

API pricing: $10 per million input tokens, $50 per million output.

Why you care: every major lab now ships a model that can do tasks while you step away from the keyboard. The difference between 'AI chat' and 'AI that actually does the job' just got a lot smaller.

On September 4, Google started removing Google Assistant from Android phones, tablets, Wear OS watches, and Android Auto. Gemini — Google's AI — now fires when you long-press the power button or say 'Hey Google'. The rollout could take a few weeks to reach all devices, but the switch is one-way: once Gemini is on, you cannot go back to Assistant.

Assistant launched in 2016 — ten years on the Android homescreen, gone in a gradual rollout.

Why you care: if you use an Android phone and rely on voice, your assistant just changed. Gemini is significantly more capable than the old Assistant — it handles multi-turn conversation, summarises emails, and acts across apps — but it is also less predictable than the command-style Assistant you may have been used to.

On September 8, Meta launched Muse — a personal AI agent that connects to your Instagram, Facebook, email, and messages and runs tasks for you around the clock. The headline is the privacy architecture: Zuckerberg recruited Moxie Marlinspike, founder of Signal, to build the confidential virtual machine underneath it. Zuckerberg's claim: "Even Meta cannot see the content that is in there."

Muse launches with up to 100 million tokens per week at no cost, with a sentinel agent that monitors incoming and outgoing data and requires your approval before taking any sensitive action.

Why you care: this is the first AI agent from a major social platform that leads with a privacy promise — and hiring Signal's founder to build it is a meaningful signal that this isn't just marketing.

We built our entire digital-products store with AI — four guides written, designed, covered, and priced in days, not months. Every one made the exact way we teach it. browse the store →

Check your Claude API usage dashboard — in early September, cache reads dropped from $1.00 to $0.25 per million tokens. If you pass a long system prompt on every call, enabling prompt caching could cut that part of your bill by 75%. Takes about ten minutes to set up in the API.