Anthropic's new small model costs a tenth of what its last one did, and that changes which jobs are worth automating. That story leads. Behind it: the Claude Code mods from this week's reel, a lawsuit over a shared AI pricing tool, and an invisible watermark heading for ChatGPT text in Europe. Six stories, each with the steps.
In today's issue
Claude's new small model costs a tenth of its last one
Ten Claude Code mods you can build by asking
McDonald's is being sued over a shared AI pricing tool
ChatGPT text will carry an invisible watermark in the EU
Claude now works inside Google Docs, Sheets and Slides
Claude feels dumber? Run these checks before you blame the model
Claude's new small model costs a tenth of its last one
Anthropic released Claude Haiku 5.5 on 7 October. It is a small model built for high-volume work like summaries, classification and live customer support, and Anthropic says it costs around 75% less to run than Haiku 4.5 on average. The same post halved the price of Sonnet 5.5 cache reads and added monthly API credits for Max and Team plans.
Why it matters
Haiku 5.5 costs $0.10 per million input tokens and $0.50 output for prompts up to 100,000 tokens. Haiku 4.5 was $1.00 and $5.00. Sonnet 5.5 is $2 and $10.
It scores 72.4% on OSWorld 2.1 (offline subset), a test of operating a real computer, against 15.7% for Haiku 4.5. Anthropic says it fits narrow jobs like compaction and summaries, and that Sonnet 5.5 (70.6% on Terminal-Bench 4.0, against 39.2% for Haiku) stays the pick for complex coding.
Sonnet 5.5 cache reads fell from $0.20 to $0.10 per million tokens, which Anthropic says cuts most agentic work by about 20%. Max 5x subscribers get $100 a month in API credit, Max 20x get $200 and Team gets up to $500 pooled, rolling out this week.
How to use it
List the repetitive text jobs you run: tagging inbound enquiries, summarising call notes, first-pass replies.
Pick one and run 20 real examples through
claude-haiku-5-5.If it gets nearly all of them right, switch that job over. Send the misses to Sonnet 5.5.
Ten Claude Code mods you can build by asking
Claude Code mods are plugins that change how it looks and behaves: a pane beside the transcript, a band above the prompt, a command that runs your own code. You describe the mod and Claude writes it. This week's reel covered three of them, and the full set of ten prompts, word for word, is here.
Why it matters
A mod can hold a tool call while it asks you a question, or send one request to a different model. That makes "send simple tasks to the cheapest model" a one-line prompt, and it pairs with the Haiku story above.
Mods are not sandboxed. A mod runs with your permissions, so build your own and read anyone else's first.
Mods need Claude Code v2.1.287 or later in the terminal.
How to use it
Run
claude --version. If it is older than 2.1.287, update.Paste: "Build me a side pane that shows me everything that I'm running." Claude builds and loads it.
Run
/pluginto confirm it loaded. If a mod misbehaves, start Claude Code with--safe-mode.Before installing someone else's mod, run
claude plugin validate ./their-modto see what it asks Claude Code to do.
McDonald's is being sued over a shared AI pricing tool
A proposed nationwide class action filed on 2 October in a Chicago federal court says McDonald's pricing platform lets independent franchisees exchange nonpublic price and sales data, which the complaint calls "algorithmic price-fixing" (Guardian). McDonald's says "AI does not set menu prices at McDonald's restaurants," and that franchisees do. These are allegations, not findings.
Why it matters
The complaint says the platform draws on data from millions of daily transactions to set menu prices across thousands of US restaurants.
At least 90 bills against algorithmic price-fixing have been filed across the US this year, according to Lindsay Owens of the Groundwork Collaborative, quoted by the Guardian.
How to use it
List every tool that suggests or sets your prices: rate cards, dynamic pricing, competitor trackers.
For each, ask the vendor in writing whether it pools nonpublic data from other businesses, yours included.
Keep a one-line note whenever you override a suggested price and why. It shows the decision was yours.
Stick to public competitor prices, and ask a solicitor if unsure. This is not legal advice.
ChatGPT text will carry an invisible watermark in the EU
To meet EU AI Act rules, OpenAI says it will add an invisible watermark to eligible ChatGPT and Codex text in the EU "over the coming weeks". API customers anywhere can opt in from now. It stays off by default in the API.
Why it matters
At a 1% false positive rate, OpenAI's detector found the watermark in about 80% of 200-token passages and about 95% of 400-token passages.
Editing weakens it. Swapping 10% of the words for synonyms cut detection from about 92% to 66%, and swapping 25% cut it to 17%.
OpenAI says a missing watermark does not prove a human wrote the text.
How to use it
Decide your disclosure line for AI-drafted copy on your site, emails and proposals, and write it down.
Treat "the detector found nothing" as no evidence either way, when you check a freelancer's work or a client's.
On the API and need to show provenance? Opt in. Otherwise leave it off.
Claude now works inside Google Docs, Sheets and Slides
Anthropic put Claude for Google Workspace into public beta on all paid Claude plans on 6 October. It opens in a sidebar next to your file, reads what you have selected and edits in place.
Why it matters
In Sheets it writes formulas, builds pivot tables and native charts, and can pull a range into Python for joins and cleaning.
In Slides it builds slides from your deck's own layouts and flags elements that overlap or run off the slide.
The default is "Ask before edits": each change waits for your approval. "Accept all edits" is optional. Admins can deploy it to a whole domain from the Google Admin console.
How to use it
Install Claude from the Google Workspace Marketplace, open a file, then go to Extensions > Claude > Open Claude.
Pick one messy sheet and ask for a budget against actuals report, one tab per team, with anything over budget flagged.
Keep "Ask before edits" on for the first week. Switch to "Accept all edits" only on jobs you can check at a glance.
Claude feels dumber? Run these checks before you blame the model
Opus 5.5's default effort is medium, Anthropic's launch post says, and that alone can feel like a downgrade if you were used to higher settings. After a week of "Claude got nerfed" posts, this week's reel ran the eight checks in order, and the full list with a weekly test prompt is on the site.
Why it matters
Anthropic says Opus 5.5 at default effort beats Opus 5 at max effort on Terminal-Bench 4.0 for about a fifth of the cost. Medium is tuned for speed and cost, not broken.
Most "dumber" days come from the first six checks: effort level, a full context window, a stale chat, a long task that needed compacting, a wrong turn that needed rewinding, or old rules in your memory file.
The test: one fixed prompt every week, same model, same effort level, scored out of 5. One bad week is noise. Two in a row at the same settings is worth acting on.
How to use it
Run
/effort status. Use/effort highfor hard work only.Run
/contextto see how full the window is, and/clearbetween unrelated tasks.Paste the weekly test prompt from the page into a fresh chat, and log the score with the date, model and effort level.
Also
ChatGPT Pro now has a $500 Pro 500 plan that includes Astra Ultrafast. Pro 200 subscribers who were active between 22 and 29 September keep the old allowance through 29 October, then move to a lower one at the same $200 (OpenAI).
Mistral Large 4 is a 1 trillion parameter model with 52 billion active, in public preview now, with open weights promised by the end of the month (Mistral).
Anthropic's startup program gives approved companies a free year of Claude Team for up to five seats plus $1,000 in API credits (Anthropic). Founded in the last 5 years or funded in the last 2.
Missed the reels? The 4 plugins I run in Claude Code every day come with install commands, and the skill that edits my reels is six stages you paste in.
Your move
Hit reply with one job you would hand to a cheap model, like sorting emails, tagging reviews or summarising calls. I will run the most common ones through Haiku 5.5 and share what held up next week.
Mike
Automate & Hustle. The AI week in five minutes, and what to do with it.