Claude’s 10 New Updates: Which Ones Change How You Build

Ten launches in one video is a lot of surface area. Duncan Rogoff’s “Claude Just Changed” walks through them cleanly, one by one, with a verdict on each. It is a good recap. What it can’t tell you is which of the ten changes your week.

I run product the way I learned to in enterprise roadmaps: a release note is a portfolio, and the job is triage. So this is not a second recap. It is my sort of the same ten launches for someone who builds alone or with a small team: which ones change how the work gets done, which ones are pleasant, which ones you can safely ignore, and what to try first. I checked every product claim against Anthropic’s own pages. Where the video and the documentation disagree, I say so.

What actually shipped, and when

The ten updates in the video landed over roughly five weeks, from early September to early October 2026, not in a single month.

That matters for one practical reason: some of these are already settled and some are days old. Opus 5.5 arrived on September 22 and has had a month of real use. Motion and the Google Workspace add-on are early betas. Claude on CarPlay was never announced at all. Duncan says as much and attributes it to news coverage, which is the honest way to handle it. Two features, Dashboards and the wrap-up allowance, have help-center pages but no launch date I could find.

One launch is missing from the ten because it came the same day the video went up: Claude Duet, Anthropic’s proactive personal agent. I covered it separately in Claude Duet: What to Delegate, What to Keep, so I won’t repeat it here.

My sort: three that matter, four that help, three that are noise

An update matters when it changes what you delegate, what it costs, or what can go wrong, rather than just where a button lives.

Here is my placement on two axes. Impact is whether the update changes how work gets done. Effort is everything between reading about it and getting value from it: setup, new habits, data plumbing, a plan upgrade, a waitlist. Drag the dots if your situation differs. A team that lives in Google Sheets should pull number 6 straight up. A team with a data warehouse should pull Dashboards left.

The top-left quadrant is where I’d spend this week. The top-right is where I’d plan a deliberate pilot with a definition of done. The bottom-left is free upside: no reason not to use it, no reason to reorganize around it. The bottom-right is what I’d leave alone until something changes.

The 5.5 models: routing is the real update

The Claude 5.5 family is three models at three price points, and the useful change is the gap between them, not any single model.

Duncan’s advice is sensible: Sonnet for everyday work, Opus for long and hard jobs, Haiku for the small things you run all day. The numbers behind it are on Anthropic’s pages. Opus 5.5 lists at $4 per million input tokens and $20 output, and Anthropic says that at default settings it will “cost 40% less than Opus 5 on typical workloads.” Note what that claim is: the per-token cut is 20%, so the rest comes from Opus 5.5 using fewer tokens per task, by Anthropic’s own testing. Sonnet 5.5 is $2 and $10. Haiku 5.5 is $0.10 and $0.50 for prompts up to 100k tokens, and Anthropic names subagent work alongside the bigger models as one of its intended uses.

That last line is the one I care about. Haiku 5.5 costs a fortieth of Opus 5.5 per token. If you run agents, a fortieth is no longer a rounding decision; it changes which jobs are worth automating at all. Summaries, classification, commit messages, log triage and the first pass of a search are the kind of work I would test there first.

The calculator prices tokens, not outcomes. A cheap answer you have to redo costs you twice: once in tokens and once in your attention. The denominator I use is the cost of work you can actually use, not the cost of a call. Route to the small model where you can check the output cheaply, and keep the expensive model where a wrong answer is expensive.

If you are on a Pro or Max subscription rather than the API, the same logic applies to your usage limits instead of your invoice. A Haiku helper burns less of the window than an Opus one.

One Claude, plus Docs, Slides and Design: the app stops asking you to choose

Anthropic merged Cowork into ordinary conversations and put document, deck and design creation into any chat, so you no longer pick a tool before you start.

The release notes date both to September 16. Cowork in every conversation is rolling out to Pro and Max. Docs, Slides and Design are available on every plan including Free, which is unusual for something this capable. Duncan says all three are out of beta; that is true for most plans, but on Enterprise they are still beta and off until an owner turns them on.

From a product point of view, the merge is the more interesting change. Asking users to classify their own task before they begin is a tax, and it is the kind of tax that quietly kills adoption in large organizations: people pick the wrong mode once, get a poor result, and stop. Removing the choice is a small design decision with an outsized effect. For a solo builder it mostly means one fewer thing to think about.

Docs, Slides and Design are where I’d expect the quickest visible return. I have built client decks by hand as HTML. A deck from notes you already have is the cheapest possible test of whether the output is good enough to send, which is also Duncan’s closing suggestion, and I agree with it.

Dashboards, Motion and Google Workspace: real, but gated

These three are substantial products whose value depends almost entirely on where your data and your team already live.

Dashboards are in beta on paid plans. According to the help center, Claude writes a SQL query for each chart, every chart shows its query and when it last refreshed, and the supported warehouses are Redshift, BigQuery, ClickHouse, Databricks and Snowflake, plus connected apps such as Salesforce. Showing the query is the right design: a chart that hides its query is the most persuasive way to be wrong. If your numbers live in spreadsheets and a payment dashboard, this is not your update yet. The video also mentions sending dashboards to Mixpanel, PostHog and Grafana; the help article only says you can send one to another tool, without naming any.

Motion builds short animations as code from your own text, charts and images, with no video generation model, so no synthetic people. The help center confirms it is a beta on Team and Enterprise only. That alone moves it to the bottom of a solo builder’s list, whatever its merits.

Claude in Google Docs, Sheets and Slides is a sidebar add-on from the Workspace Marketplace, in beta on Pro, Max, Team and Enterprise. The help article describes an “Ask before edits” default and changes that appear in normal version history under your name. For a team that runs on Google files this is the most practical update of the ten. For me it is close to irrelevant, which is exactly why the radar is draggable.

Projects: parallel threads toward one goal, and a bill to match

Redesigned Projects turn a Claude Code project from a folder of chats into a goal that a coordinator splits into parallel threads, each on its own branch and copy of the repository.

This is the update closest to how I already work. On my game project, Ballistica, I run several agent lanes in parallel and merge what comes back, and most of the effort is coordination: who touches what, what was decided, what conflicts. Anthropic’s announcement describes exactly that structure as a product, with a coordinator that scopes, delegates, reviews and assembles, shared memory for decisions, and overlapping edits that surface as merge conflicts rather than silent overwrites.

Two corrections to the video. Duncan says every Pro and Max user can get it now and that it runs in the cloud or on your own computer. The announcement says the beta started with select Pro and Max subscribers using cloud sessions, with a waitlist, and that running on your own machine is coming soon. And the announcement is candid about cost: several full Claude Code sessions at once reach usage limits faster.

The product-leader caution is the one I wrote about in More Agents. One Person Who Signs Off. Five threads finishing at once is five reviews landing at once. Parallel execution only helps if your review keeps up.

Mods: your harness becomes yours

A mod is a plugin of JavaScript or TypeScript event handlers that run inside Claude Code and can observe, rewrite or answer events such as tool calls, prompts and screen drawing.

This is the one Duncan saved for last, and I’d have done the same. He frames it as a browser extension for Claude Code and shows the fun ones: Doom while Claude works, Tetris above the prompt, a breathing animation. He also passes on Boris Cherny’s point that people work differently, so their Claude Code should not have to be identical. The documentation is more sober and more interesting. A hook runs before Claude Code acts on an event. It can let the event through, change it, or handle it itself so the default never runs. Anthropic even moved its own /diff into a mod.

Here is why I put mods in the top-left of the radar. My CLAUDE.md has a line that says to confirm before any force push, because a bidirectional deploy sync depends on that history. Today that line is a request Claude reads. A mod like Anthropic’s blast-radius sample turns it into a gate the harness enforces. That is the difference between a policy and a control, and it is the same shift I described in the Claude Code agent harness: reliability comes from the system around the model, not from the model remembering your preferences.

The warning matters as much as the feature. Mods run with your permissions and are not sandboxed. They can read your environment variables, approve a tool call before you are asked, and spend your usage. Install them like you would install a dependency with production access: from authors you trust, after reading what the docs’ claude plugin validate command lists about them.

The wrap-up allowance: the smallest fix on the list

The wrap-up allowance lets Claude Code finish the step it is on when you hit a five-hour usage limit mid-response, charged to your weekly limit.

There is nothing to adopt here, which is why it sits at zero effort. Anthropic’s help article adds the detail the video skips: Pro gets it up to once per weekly usage period, Max each time you hit a five-hour limit within the weekly limit, it does not cover new messages, and it is rolling out gradually. The value is avoiding the worst kind of failure, a file left half-edited with no clean point to resume from.

Which updates matter for you

The right first update depends on your role and your plan more than on the update itself, because four of the ten are restricted by plan.

Pick the closest role and your actual plan. The list re-ranks, greys out what you can’t use, and gives you a three-line plan to copy.

My first three Claude updates to try (Solo builder, Max plan):
1. Opus, Sonnet & Haiku 5.5: Sonnet by default, Haiku for bulk and helper work, Opus for long unattended runs.
2. Claude Code mods: Turn the rules in your CLAUDE.md into gates the harness enforces.
3. Projects, redesigned: Parallel threads toward one goal. Powerful, and it spends usage fast.

CarPlay is the clearest example of an update that is pleasant and irrelevant to the work. It is voice-only, it can’t control your car or phone, and you open the app each time. If you think out loud on a commute, it is nice. It will not change your output.

What I’d try first, in one week

The best adoption plan for a batch of updates is one bounded test per update that matters, each with a result you could fail.

1. Day 1: one deck from existing notes. Use Slides on a real meeting. Done means you’d present it with only light edits.
2. Day 2: route one recurring job to Haiku 5.5. Pick something you can check at a glance, such as summaries or labels. Done means you accept the output at the same rate as before.
3. Day 3: one guard mod. Ask Claude Code to write a mod that holds the one command you would hate to run by accident. Read it before you turn it on.
4. Day 4: one goal for Projects, if you have access. Choose work that genuinely splits by repository or area. Watch your usage meter, not just the threads.
5. Day 5: decide. Keep what saved review time; drop what produced more output to check.

A launch matters when it changes what you can safely hand off. Most of the ten change where you click. Three change what you delegate.

Frequently asked questions

Which Claude 5.5 model should I use by default?

Sonnet 5.5 for everyday work, as both Duncan and Anthropic’s positioning suggest. It lists at $2 per million input tokens and $10 output. Move long, unattended or high-stakes runs to Opus 5.5, and repetitive, checkable jobs such as summaries, labeling and helper subagents to Haiku 5.5, which is a fortieth of Opus’s per-token price. Judge each choice by accepted output, not by the price of a single call.

Are Claude Code mods safe to install?

Only as safe as their author. Anthropic’s documentation says mods run with your permissions and are not sandboxed: they can read files and environment variables, make network requests, approve tool calls and spend your usage. Install from people and marketplaces you trust, inspect what a mod hooks with claude plugin validate, and turn mods off with --safe-mode if a session behaves oddly.

Can every Pro and Max user use the redesigned Projects today?

Not according to Anthropic’s September 17 announcement. The beta began with select Pro and Max subscribers using Claude Code cloud sessions, expanding over the following week, with a waitlist for others and Team and Enterprise later. Threads run in the cloud; running on your own machine is described as coming soon. Expect it to use your limits faster, because each thread is a full session.

Is Claude Motion available on Pro or Max?

No. Anthropic’s help center lists Motion as a beta on Team and Enterprise plans, on by default for Team and off by default for Enterprise until an owner enables it. It animates your own text, charts, shapes and images as code rather than generating video footage, so you can edit numbers and timing after the fact and export when you are happy.

Do I need to do anything to get the wrap-up allowance?

No. It is rolling out gradually to Claude Code on Pro and Max. When a five-hour limit lands mid-response, you may see “Usage limit reached · wrapping up” while Claude finishes the current step. Pro gets it up to once per weekly usage period; Max each time, within the weekly limit. The extra usage counts against your weekly limit, and new messages are not covered.

Sources and limits

Written October 22, 2026, the day after the video was published. The ten-update framing, the examples of community mods and the Boris Cherny remark come from Duncan Rogoff’s video. Dates, prices, plan availability and feature behavior come from Anthropic’s release notes, model pages, announcement and help center, linked inline. The CarPlay launch and the dates for Mods, Motion and the Google Workspace add-on come from news coverage, as the timeline marks. Placements on the radar, fit scores and the one-week plan are my editorial judgment, not measurements. Beta features change; check the linked pages before you buy a plan for one.