Nile Automate
← All issues

This Week in Claude · 2 October 2026

This Week in Claude: 2 October 2026

Claude Sonnet 5.5 arrived at the same price, faster and up to 30% cheaper per task, and Sonnet 4.5 got a retirement date of 30 November. Re-pin everything.

Week of 25 September to 2 October 2026. Scan it in 30 seconds, read the takeaways in three minutes, or go the whole way in ten.

Nhial Majok
Nhial MajokFounder · AI & Automation Lead · Melbourne

The scan

Sorted by what to do about it.

  • Act now

    Anthropic shipped Claude Sonnet 5.5 on 28 September at US$2 in and US$10 out per million tokens, 30% faster and up to 30% cheaper per task than Sonnet 5, then deprecated Sonnet 4.5 two days later with API retirement on 30 November. Anything pinned to Sonnet 4.5 needs a migration owner.

  • Act now

    Notion renamed its MCP tool notion-query-database-view to notion-query-data-sources on 29 September. Any Cowork skill or script that calls the old name fails now.

  • Watch

    HubSpot shipped Claude for Small Business workflows inside Cowork on 28 September, then added a Xero connector, per-connector tool permissions and a default 50% cap on AI credit spend.

  • Watch

    Anthropic's plugin submission portal opened on 25 September for paid plans, so a plugin can be published once and installed across Claude, Claude Code and Cowork.

  • Noise, for now

    Anthropic's leaked IPO filing (US$4.59 billion of 2025 revenue, US$518 billion of compute commitments, a mid-November listing at US$1.8 to 2 trillion). Big numbers. Nothing to do on Monday.

Numbers that moved

  • US$2 and US$10 per million tokens for Sonnet 5.5
  • 30% faster than Sonnet 5
  • 30 November for Sonnet 4.5 API retirement
  • 70.6% on Terminal-Bench 4.0
  • 50% default cap on HubSpot AI credit spend.

Takeaways for your business

What it means on Monday.

  1. 01Sonnet 5.5 arrived and Sonnet 4.5 got a retirement date. Re-pin everything.

    What happened
    Anthropic released Claude Sonnet 5.5 on 28 September, model ID claude-sonnet-5-5, at US$2 in, US$10 out and US$0.20 per million for cache reads, unchanged from Sonnet 5, and says it runs 30% faster and costs up to 30% less per task. It scores 70.6% on Terminal-Bench 4.0, above Opus 5.5's 66.4%. Two days later Anthropic deprecated Sonnet 4.5 with retirement from the API on 30 November. Haiku 5.5 is still "in the coming weeks" and Haiku 4.5's retirement stays at not before 15 October.
    Why it matters
    For most small-business workflows the mid-tier is now the default, and it just got faster at the same price. The clock that matters is 30 November: any Zapier, Make, n8n or custom script pinned to claude-sonnet-4-5-20250929 stops working that day.
    Next 30 days
    List every automation with a model ID in it. Move anything on Sonnet 4.5 to Sonnet 5.5 and test it, because the new model rejects a few parameters the old one accepted. Hold the Haiku moves until Haiku 5.5 ships. Ask any vendor that bills you for "AI features" which model they're on and what they did with the price drop.
    For operators

    Sonnet 5.5 has five breaking changes from Sonnet 5. Up-front thinking is turned off with thinking: {"type": "between_tools"}, not "disabled", and only at high effort or below. Forced tool use (tool_choice: any or tool) returns a 400. Thinking blocks are now account-bound, so a transcript replayed from a different Anthropic account silently drops them. The computer_20251124 tool is rejected. Non-default temperature, top_p and top_k return 400.

  2. 02Always-on Claude workflows are being sold to small businesses. Trial one. Govern it like a staff member.

    What happened
    HubSpot shipped Claude for Small Business workflows inside Cowork on 28 September (/monday-brief, /speed-to-lead, /crm-autopilot), then added a Xero connector, per-connector tool permissions and a default 50% cap on AI credit spend in its September release.
    Why it matters
    The pitch has moved from "chat with AI" to "an AI that works while you sleep", and it's being pitched at a 12-person business, not a bank. It ships with approval gates and a spend cap. The gate is the feature. The risk is the owner who turns it off in week two because approving things is annoying.
    Next 30 days
    If your business already lives in HubSpot, run it with every approval on for a fortnight and count how many approvals you'd have been happy to skip. That count is your allow list. Set the spend cap before you set anything else.
    Say it on a call
    "Keep the approvals on for two weeks, and we'll turn off only the ones you'd have said yes to every time."
    For operators

    The vendor spend meters are new and not all reliable. Claude Code 2.1.286 fixed its gateway spend meter under-pricing one-hour cache writes at the five-minute rate, and HubSpot's cap defaults to 50% of account credits. Put your own usage logging in front of every agent rather than trusting the vendor dashboard. Anthropic's plugin submission portal opened on 25 September for paid plans, so a Xero, ServiceM8 or Deputy plugin can now be published once and installed by clients across Claude, Claude Code and Cowork instead of hand-installed.

  3. 03Your Notion automations may already be broken.

    What happened
    Notion renamed its MCP tool notion-query-database-view to notion-query-data-sources on 29 September and raised MCP rate limits.
    Why it matters
    Small, boring and the kind of thing that breaks an automation at 3am. If any Cowork skill or script calls the old Notion tool name, it fails now.
    Next 30 days
    Search every skill and prompt file for notion-query-database-view and replace it.

What I’m doing about it

Moving the two client workflows still on Sonnet 4.5 to Sonnet 5.5 before the 30 November retirement.

Watchlist

What we’re tracking.

  • Anthropic IPO. Updated. The S-1 leaked on 28 September (2025 revenue US$4.59 billion, operating loss US$8.06 billion, US$518 billion of ten-year compute commitments, 80% non-cancellable) and Bloomberg reports marketing from the week of 9 November with trading before 26 November at US$1.8 to 2 trillion. Still no public filing on EDGAR.
  • Claude 5.5 rollout and 4.5 retirements. Updated. Sonnet 5.5 shipped 28 September. Sonnet 4.5 deprecated 30 September, API retirement 30 November. Haiku 5.5 still "in the coming weeks". Haiku 4.5 retirement not before 15 October, no notice yet.

Corrections from last week: none.

The full report

Anthropic and Claude

So what: a new workhorse model at the old price, and an end date for the one many automations still run on.

Claude Sonnet 5.5 (confirmed, primary source). Launched 28 September, US$2 in, US$10 out, cache reads US$0.20, five-minute cache writes US$2.50, batch at 50% off, 1M context, 128K max output, knowledge cutoff June 2026. Anthropic claims 30%-plus faster and up to 30% cheaper per task than Sonnet 5. Benchmarks (Sonnet 5.5 / Sonnet 5 / Opus 5.5): Terminal-Bench 4.0 70.6% / 10.3% / 66.4%; OSWorld 2.1 80.1% / 57.0% / 81.8%; GDPval-AA v2.1 1844 / 1449 / 1846. Retirement not before 28 September 2027. Available on AWS, Google Cloud and Azure. Haiku 5.5 "in the coming weeks".

Also this week. Claude for Government went GA on 30 September with fixed-commitment pricing and usage caps, Claude for Microsoft 365 entered early access (reported, TechRadar), and Barclays committed to Claude Code for 50% of its developers by end 2026 (confirmed).

Deals and the IPO. Anthropic and Akamai signed a US$11.6 billion, seven-year CPU-inference deal, expandable to about US$20 billion, with a warrant for up to 5% of Akamai (confirmed, 25 September). S-1 details leaked 28 September (reported, Reuters and FT), and Bloomberg reports a mid-November listing at US$1.8 to 2 trillion (reported, 1 October). The S-1 shows more than US$100 billion of Anthropic's commitments sit with AWS (reported).

For operators: Claude and agentic tooling notes

So what: a model migration and a new way to ship plugins. Read the release notes before Monday.

All confirmed from Anthropic release notes and the Claude Code changelog unless marked. Sonnet 5.5 breaking changes are in takeaway 1. Sonnet 4.5 (claude-sonnet-4-5-20250929) is deprecated as of 30 September and retires 30 November; the migration guide points to claude-sonnet-5-5. Claude Code 2.1.287 (1 October) introduced Mods, TypeScript functions inside plugins that can rewrite prompts and replace built-in features; they are not sandboxed and have full machine access, and Team and Enterprise get a sec-default mod that loads first. The same release defaults Opus 4.7-plus and Fable to 1M context on Bedrock, Vertex and Foundry (CLAUDE_CODE_DISABLE_1M_CONTEXT=1 keeps 200K), supports MCP sign-in prompts on the 2025-11-25 protocol, and adds prompt_text to the OpenTelemetry user_prompt event, so mask it wherever you mask prompt. Claude Code 2.1.286 (30 September) caps failed-request retries at 14, auto-runs a project skill named verify before commits, and fixed the apps-gateway spend meter. The plugin submission portal opened 25 September for paid Claude plans, with safety scanning and install analytics. Notion's MCP renamed notion-query-database-view to notion-query-data-sources (29 September, reported, corroborated by the live tool list) and raised limits to 20 search and 20 query calls per 10 seconds. MCP spec: no new release; current is 2026-07-28. Retirements to diarise: Claude Haiku 4.5 not before 15 October, Claude Sonnet 4.5 from the API 30 November.

Sources (17)

If something here changes your Monday

Let’s work out what to do about it.

Get notes like this in your inbox when we share them.

You join the Field Notes list. Unsubscribe in one click.

Rather have someone set it up with you than work it out alone?

Book an intro call

Free, 30 minutes. No obligation.