What changed
Anthropic released Claude Sonnet 5.5 on September 28, 2026. It costs exactly what Sonnet 5 cost: $2 per million input tokens, $10 per million output tokens. But it finishes tasks with 30% fewer tokens, so the same dollar bill goes further. It scored 70.6% on Terminal-Bench 4.0, a test of whether AI can work inside a command line. That beat GPT-6 Astra's 57.9%. The model is now the default on Claude's free tier at claude.ai, replacing the older Sonnet. OpenAI's free tier still uses Luna 5.6, so Anthropic now offers a stronger model for free.
What you can do with it
Pick one task you've been paying someone to do every week: the meeting summary, the customer support answer, the invoice description, the blog post draft. Run it through Claude Sonnet 5.5 at claude.ai for free first. If you hit the free limit — about 15 to 40 messages every five hours — and the output is good enough, calculate what the API would cost. At $2 per million input tokens, one thousand 500-word documents cost about one dollar to process. Compare that to what you pay now. If you're spending $200 monthly on a task AI can handle, try routing half the volume to Claude for a month and compare the results.
What to keep in your own hands
Don't paste in customer data, employee records, or financial details until you've read Anthropic's data policy and confirmed whether your account has zero data retention enabled. The API offers zero retention; the free web app does not by default. Don't use AI to write anything a client or regulator will hold you responsible for without a person reading it first. Don't assume cheaper per token means cheaper per month — if the model is easier to use, you might use it more. Track your spending weekly, not monthly. And keep one person on your team who knows how to do the task without AI, so you're never stuck when the service is down.