Simon Willison just shipped ttok 1.0, a token-counting CLI tool that does one thing well: tell you exactly how many tokens a piece of text will cost you in OpenAI’s API.

The tool hasn’t been updated in two years. Version 0.4 dropped on October 8th with some housekeeping: fixed a Click warning, updated CI, added a --list-models flag. Standard maintenance stuff.

Then Willison ran uv tool upgrade ttok, piped a file through it, and realized it was still defaulting to the GPT-4 tokenizer. In 2026. When GPT-5 and GPT-6 are the current models.

Version 1.0 shipped the next day with the fix.

Why this matters

Token counting is more important than it used to be. Context windows are massive now, which means you can easily blow through your budget if you’re not paying attention. Knowing your token count before you make an API call is the difference between “this costs $0.12” and “this costs $12.”

ttok uses OpenAI’s open source tiktoken library under the hood, so it’s accurate. Run it with uvx and you don’t even need to install it:

cat file.txt | uvx ttok

That’s it. You get a number back.

The GPT-6 tokenizer question

There’s a wrinkle here. OpenAI hasn’t officially confirmed that GPT-6 uses the same tokenizer as GPT-5. There’s an angry GitHub issue about it. But Willison found a commit from William Liu reporting on an experiment that suggests they’re the same, so he went with it.

If you’re counting tokens for GPT-4 or earlier models, you’ll need to specify with --model. The default is now GPT-5/6.

Who should care

If you’re building anything that calls OpenAI’s API, you probably want this. It’s especially useful if you’re doing RAG, summarization, or anything else where you’re programmatically stuffing text into prompts.

If you’re just using ChatGPT in the browser, you don’t need it.

The tool is on PyPI and works with pip, pipx, or uvx. It’s open source, MIT licensed, and the code is straightforward enough that you could fork it if OpenAI ever changes tokenizers again and Willison takes another two-year break.

Other stuff that shipped this week

Goodfire launched inside-out AI monitoring. Instead of running a second AI model to watch your AI agents (expensive), their system hooks into the model’s internals and only escalates when something looks wrong. They claim it’s cheaper. No pricing details yet, so we’ll see.

Google added agentic AI to Gemini for businesses. The agent can delegate to subagents, use multiple models, and gets its own workplace identity with an email address. It’s enterprise-only for now. If you’re using Google Workspace and want your AI to file TPS reports, this is for you.

Sophos cut threat investigation time by 96% using OpenAI Daybreak. They’re automating 52% of MDR cases while keeping humans in the loop. Daybreak is OpenAI’s reasoning model for specialized tasks. If you’re in cybersecurity and drowning in alerts, worth a look.

Anthropic updated their usage policy. You can’t repeatedly abuse Claude anymore (repeatedly being the key word, normal frustration is fine). They also banned election interference, deceptive campaigns, weapons software, and surveillance. The policy now explicitly covers things that were probably already violations but are now spelled out.

The ttok release is the most useful thing here for most developers. Small tools that solve one problem well are underrated.