AI news this week for small business: GPT-6 for everyone, cheaper Claude and spending caps (Oct. 12, 2026)

01The week at a glance

021. GPT-6 arrives in ChatGPT for almost everyone (October 7)

On October 7, OpenAI started rolling out GPT-6 with Intelligent UI in ChatGPT's Chat tab, globally: Plus, Pro, Business and Enterprise first, then Free and Go from the next day. Answers can now mix text, visuals and interactive elements, such as side-by-side comparisons, diagrams, and calculators you use right in the conversation. Enterprise access depends on admin settings, and the models behind ChatGPT Work and Codex don't change.

What it changes for a small business: you can ask for a tool, not just an answer. A window-cleaning company could ask for a small estimator built on its own price per window. The answer becomes something you click, not only something you read.

032. Claude Haiku 5.5 cuts the price of repetitive AI tasks (October 7)

Anthropic released Claude Haiku 5.5 on October 7, a small model for high-volume work like summaries, classification and live support. For prompts up to 100,000 tokens it costs US$0.10 per million input tokens and US$0.50 per million output tokens, against US$1.00 and US$5.00 for Haiku 4.5. The same day, Anthropic halved Sonnet 5.5's cache-read price and announced a monthly API credit for Max and Team subscribers (up to US$500 a month for Team, pooled across users). Canada is on Anthropic's list of supported countries.

These are model list prices only, before any automation platform's fees. The point stands: the model is no longer the expensive part. Your team's time checking the results is.

043. AI by Zapier: approvals per tool and a 75-task brake (October 6)

On October 6, Zapier moved everything Zapier Agents could do into AI by Zapier, one AI step inside an ordinary Zap. You can now require approval before running, set per tool; Zapier notes agents used to be all-or-nothing. AI steps bill as tasks: Zapier says default models are free and other tiers carry multipliers, so a step can cost 1, 3 or 5 tasks. And if a single run crosses 75 tasks, it pauses for a human instead of looping quietly. Some capabilities are still rolling out; Canada isn't mentioned.

What it changes: the AI can read a spreadsheet or an inbox on its own, but has to ask before it emails a client or updates your CRM. That's the difference between an assistant and a liability.

054. Spending caps are becoming normal (October 8)

On October 8, Google announced the Gemini agent: one agent for questions, documents, media and code that keeps working in the cloud after you close your laptop. The most useful detail for a small business: a hard spending cap per project, set in the Cloud Billing Console. When the cap is hit, that project's agent pauses until you resume it. The announcement gives no general price, plan or date, and doesn't mention Canada; only financial-services and legal versions are in preview.

The same day, Linear added AI spend limits per workspace, user or loop: at the limit, extra AI usage pauses until it resets or an admin raises it. And OpenAI went the other way: an Ultrafast mode for GPT-6.1 Sol at US$12 and US$60 per million input and output tokens, six times the standard US$2 and US$10, to cut the wait between words. Its processing options are global, US or EU.

065. A study worth reading if you have junior staff (October 7)

On October 7, Google researchers David Autor and Tanya Rodchenko described a three-month randomized trial with 133 lawyers at eleven intellectual property firms. With an AI drafting assistant, the quality of their patent drafts went up. Then everyone corrected a flawed patent without AI. Lawyers who had used the tool did better than the control group, but the gain came entirely from senior lawyers. Juniors, under seven years of experience, showed no discernible improvement on average.

Improving performance and building durable professional judgment are different goals.

Google Research, October 7, 2026

What it changes: it's one profession over three months, and the authors say so. But if your juniors draft everything with AI, keep part of their review work AI-free, and ask them to fix what they find, not just point at it.

07Before you build on any of it: the six-line announcement check

Announced isn't the same as available to you. Here is the prompt we use on any AI announcement. Paste the vendor's own text; the rule that matters is the third sentence.

You help a [YOUR BUSINESS: size, trade, province] decide whether an AI announcement changes anything for them. Use ONLY the announcement text below. Never guess: if the text doesn't say it, write NOT STATED.

Answer in exactly these six numbered lines:
1. What's new: one plain sentence.
2. Who can use it today (plan, preview or beta): quote the text, or NOT STATED.
3. Available in Canada: quote the text, or NOT STATED.
4. Cost: quote the text, or NOT STATED.
5. What could go wrong for a small business: one line.
6. One safe test we could run this week with fake data, or WAIT if we can't use it yet.

Announcement:
"""
[PASTE THE ANNOUNCEMENT FROM THE VENDOR'S OWN PAGE]
"""

We ran it with Claude (Anthropic) on excerpts of Google's October 8 Gemini agent announcement, for a fictional 12-person Canadian business. Lines 2, 3, 4 and 6 of its answer, as produced:

2. Who can use it today (plan, preview or beta): For the general agent, NOT STATED. The text names no plan, preview or beta. Only the industry versions have a date: "It is now in preview for Financial Services and Legal, and coming soon to Government, Healthcare, and Retail."
3. Available in Canada: NOT STATED.
4. Cost: NOT STATED. The text gives no price. It only says "Gemini monitors token usage and sandbox costs" and that you can set "a hard limit on a project's AI spend in the Cloud Billing Console."
6. One safe test we could run this week with fake data: WAIT. The text doesn't say the general agent is available to buy, available in Canada, or what it costs. Only Financial Services and Legal are in preview.

Not a bad announcement; a reminder that a keynote isn't a price list. On Zapier's announcement, the same check quoted the task costs and suggested a safe test with fake customers.

08Three more from the week

• October 6, audio uploads in ChatGPT: paid plans can turn a meeting recording into a transcript, a summary or a follow-up email draft. Transcripts may contain errors.
• October 8, failure alerts in Make: real-time notifications when a scenario run fails, on Teams (up to three subscriptions) and Enterprise plans.
• October 9, Office in the Copilot app: Word, Excel and PowerPoint editing inside Microsoft's Copilot app, first for customers in its Frontier program.

09One thing to try this week: give one AI task a brake

1. 1
   List what runs without you

   Automations, shared chat accounts, agents: anything AI does while nobody watches.

2. 2
   Pick the one that touches clients or money

   An email that goes out, a quote, an invoice, a CRM update.

3. 3
   Add one brake

   An approval before anything is sent, a spending cap, or an alert when it fails.

4. 4
   Write it down

   Add it to your one-page AI policy so the next person knows it's there.

And before you cap anything, know what the work you're automating really costs: our free ROI calculator shows it live, with every assumption visible. Its default example, 14 hours a week of quotes, email and invoicing, frees up about 360 hours a year in the typical scenario.