What's new in Claude

If you are away from your desk, you can now kick off a session from the Claude app on your phone and have it run on your own laptop. Nothing moves to the phone — your files, your folders and the work itself all stay on your machine, and the phone is just the remote control you press the buttons with.

Your laptop has to be on and reachable for this to work. It is a way to get something started, or to check in on it, without opening the lid.

Some people find Claude’s explanations helpful and some find them long. A concise output style now exists for the second group: pick it and the replies get shorter and less conversational.

It only changes the talking, not the doing. The same steps are taken, the same files are changed, the same diff is shown — you just read less commentary around it.

Every plan has a usage limit, and hitting it in the middle of something used to mean the task simply stopped and stayed stopped. Claude Desktop now waits out the reset and carries on from where it left off.

The practical effect is that a long job you started before running out no longer needs you to come back and restart it by hand.

The introductory price Sonnet 5 launched with — $2 and $10 per million tokens — is now the permanent price. A rise to $3 and $15, which had been set for 1 September 2026, was called off.

Those figures apply to people billed per token through the API. If you pay a monthly Claude subscription, nothing about your bill or your plan changes.

Opus 4.1 has been retired. If it was the model you had chosen and kept, you were moved to a current model rather than left with a dead setting.

Retirements are routine — older models are stood down as newer ones take over. The visible sign is simply that a name you recognised is no longer in the model list.

A permission mode decides how often Claude stops to ask before it does something. Auto mode — where it does not pause at every step — is now the setting a new install starts on, so this is what someone opening Claude Desktop for the first time sees. In place of those prompts, a safety check runs quietly in the background and blocks actions it judges risky.

Two things are worth knowing before that unsettles you. You can still switch back to approving each step, in the permission settings, and plenty of people do while they are finding their feet. And a diff — the side-by-side view of exactly what is about to change in a file — is still shown before anything is saved, in auto mode as much as in any other. Fewer interruptions, in other words, not less sight of what is happening.

The short version for anyone using the app: the model behind it got better, and there is nothing for you to do. Opus 5 became the default Opus in Claude Code within the same week it launched, so you were moved onto it without touching a setting.

Anthropic calls it a step-change over Opus 4.8. It reads up to a million tokens of context at once — roughly, how much material it can hold in mind in one go — and thinks before answering by default. Pricing, which matters only if you are billed per token through the API, is $5 and $25 per million tokens.

A screen reader mode was added. A screen reader is software that reads what is on screen aloud, and this mode changes how Claude presents its output so that reading it that way makes sense.

It is a small line in a release note and a large difference to the people who need it, which is why it gets its own entry here.

What Claude remembers about you used to be one long, opaque block of text. It is now a categorised list of separate entries, each of which you can read and delete on its own.

The practical effect is that memory stopped being something you either trusted wholesale or switched off. If one stored detail is wrong or stale, you can remove that detail and leave the rest.

Until this change, connecting Microsoft 365 let Claude look at your mail, calendar and files. It can now act on them: drafting and sending email, managing calendar entries, and writing files into OneDrive and SharePoint. For analysts and operations work, that removes a lot of copying between windows.

The obvious caution follows from the same sentence. Writing means it can act, and an action in a live mailbox is harder to take back than a bad summary, so read what a connector is permitted to do before you attach it to a real account.

The Desktop app now has a working browser built into it. Claude can point it at an external website or at something you are running on your own machine, look at what is rendered there, and adjust based on what it sees.

The word that matters is see. Before this, describing a page to Claude was your job; now it can check the result itself and go round again.

/doctor became a complete checkup of your setup. Run it and it looks over the installation, finds the part that is missing or misconfigured, and says what to do about it.

This is one of the more useful things here for anyone new. The hardest moment with a new tool is the one where it simply does not work and nothing explains why, and this is aimed squarely at that moment.

Administrators can now restrict the set of models available to their organisation.

This is worth knowing even if you never touch an admin screen, because it explains something otherwise puzzling: if your model dropdown is shorter than the one a friend has at another company, that is usually a decision someone made for your account, not a fault on your machine.

Claude Fable 5 became available again on 1 July 2026. It had been suspended since 12 June, three weeks earlier. Anthropic set out its reasoning for redeploying the model in a published statement.

The suspension itself is covered in its own entry, dated 12 June 2026.

Claude Sonnet 5 launched and immediately became the default model in Claude Code. If you did not change your model setting, you were simply working with the newer one — the engine behind the app got better and there was nothing to do about it.

It carries a context window of a million tokens, meaning it can hold a great deal of material in mind at once. Priced at $2 per million tokens in and $10 per million out, which is a number for people billed by the API rather than for anyone on a subscription.

Claude in Chrome became generally available. “General availability” is the point at which something stops being a trial and is treated as finished and supported.

Claude Desktop came to Linux in beta, installable on Ubuntu and Debian through apt or a .deb package.

This one matters here more than most. Until now the Desktop app ran only on macOS and Windows, which is the whole reason readers on Linux were pointed at the terminal instead. Linux readers now have a graphical option too. It is in beta, so expect rough edges and expect it to change.

If a piece of work is already being discussed in Slack, you can hand it to Claude by tagging it in the thread. It works from that thread’s context, so you do not have to re-explain what the conversation was about.

This is useful when the request and the background are already written down by colleagues — the ask stays where the discussion happened.

Wiping a conversation used to be a one-way door. After this change, /rewind — the command that steps a session back to an earlier point — can resume from before a /clear, so a conversation you cleared by mistake is recoverable.

That matters if you are still finding your footing, because clearing is one of the easiest things to do by accident. Undoing is the safety net the whole Start path leans on.

Output from a session can be published as an Artifact: a live page with its own link. The person you send it to just opens the link — nothing to install, no account setup, no copy-pasting the result into an email.

For anyone whose work ends in showing something to someone else, this is the shortest path from a finished result to a colleague looking at it.

Anthropic retired Claude Sonnet 4 and Claude Opus 4. Retiring a model means it stops being available; the names drop out of circulation and newer models carry the work.

There is nothing to do about it. If either of those names is familiar to you from an older article or an internal note, it is out of date.

Small fixes to a draft used to mean writing another request and hoping the rest survived intact. Now you can edit the draft in place — change the wording you disagree with and leave everything else alone.

For writing-heavy work, this is the difference between a tool that gives you a starting point and one you can actually finish a document in.

Three days after Claude Fable 5 launched, Anthropic suspended access to it. Access was restored on 1 July 2026.

Claude Fable 5 arrived as Anthropic’s most capable widely released model. It carries a one-million-token context window — roughly, how much text it can hold in mind at once — and adaptive thinking is always on, meaning the model decides for itself how long to think about a request rather than waiting to be told.

Pricing was $10 per million tokens in and $50 per million out, which is a number for people paying by API rather than for anyone using a Desktop subscription.

Claude Opus 4.8 was released. Anyone on the Max plan using Claude Code was defaulted to fast mode on it.

For most readers this was a quiet change: the model behind the app improved, and there was nothing to switch on.

Auto mode had been a Max-only feature. From this date it was available on the Pro plan too.

If you are on Pro and had read about auto mode without being able to find it, this is when it appeared for you.

/usage shows what is actually driving your usage limits — the allowance your plan gives you over a set window of time.

It answers the question almost every new reader asks eventually: why did I run out? Rather than guessing, you can look.

If you have asked Claude to do more than one thing at a time, keeping track of which of those sessions is still going used to be your job. Agent view gathers them into a single screen, so you can see at a glance everything that is currently running.

For most people the takeaway is the usual one: the model behind the app got better and there is nothing for you to do. Images are now read at higher resolution, so screenshots and photos come through in more detail.

The one wrinkle worth knowing is the tokenizer — the thing that chops your text into the units Claude counts. claude-opus-4-7 counts about 30% more of those units for the same words, which means usage numbers from before and after the change are not comparable with each other.

A long job used to mean sitting there, because Claude would pause for a decision and there was no way to know unless you were looking at the screen. With push notifications, you can start something, walk away, and get a nudge on your phone at the moment it needs you.

Cowork is a different tab from Code, which is the one this site teaches. Where Code is a conversation you steer as it goes, Cowork is for handing a task over and letting it run by itself in the background.

General availability just means it is no longer an experiment: on macOS and Windows it is a normal part of the app.

Typing /powerup opens a hands-on introduction built and shipped by Anthropic inside the app. Worth saying plainly: Anthropic now explains some of this material in-product, and there is no reason not to use both that and this site.

Until this point, Claude paused and asked for your approval before each step it took. Auto mode was the alternative: you hand over the wheel and it keeps going without checking in every time.

It was an option you chose here, not the way things worked out of the box. That came later — this is where the change of 3 August 2026, which made auto mode the default for everyone, starts.

This is one of the most directly useful changes on this page if your job involves numbers or explaining how something works. Ask about your figures and you can get a chart back; ask how a process fits together and you can get a diagram. You no longer have to take Claude’s written answer and rebuild it by hand in a spreadsheet or a drawing tool before anyone else can look at it.

For an analyst, that is the gap between an answer and something you can put in front of a manager. For marketing or ops, it is the same gap between a description of a funnel or a handover and a picture of one.

An add-in is Claude running inside the Office app itself rather than in a separate window. Until this date the Excel add-in and the PowerPoint add-in were two strangers: whatever you worked through in the spreadsheet had to be explained again from scratch when you opened the deck.

That is the part that changed. Work through your numbers in Excel, move to PowerPoint, and the same conversation is still there — the caveats you agreed on, the figures you decided to use, the framing you settled on. For anyone whose week is analysis on Monday and a deck on Wednesday, it removes the most tedious repetition in the whole routine.

Memory is the feature that lets Claude hold on to things about you from one conversation to the next, instead of meeting you fresh every time. From this date it was available to all users, including the free tier.

If you are not paying for anything, this is probably the most noticeable change on the whole page for you.

Cowork is a separate tab in the Desktop app from the Code tab this site is mostly about. Before this change, work in it started when you started it. After it, you can put something on a timetable and it runs on schedule.

The obvious use is anything you currently do because it is Monday, or because it is the end of the month — the recurring chore rather than the one-off question.

The context window is how much Claude can hold in mind at once within a single conversation — your documents, the back-and-forth, everything. Sonnet 4.6 shipped with a one-million-token window, offered in beta.

As with every model release, there was nothing for you to do. The model underneath the app improved, and long pieces of work got less likely to run out of room.

Thinking is the extra working-out a model does before it answers. Up to this point, how much of it happened was something you set. Opus 4.6 introduced adaptive thinking, where the model makes that call itself — more on a hard question, less on an easy one.

For you, that means one fewer setting to understand and get wrong. The model behind the app got better; there was nothing to install and nothing to switch on.

An add-in runs inside the Office app you already have open. So for deck work, this is the difference between describing your slides to Claude somewhere else and having it there in the file with you.

If building presentations is a regular part of your week, this is the entry on the page worth acting on.

Cowork is a tab in the Desktop app, separate from the Code tab this site concentrates on. From this date it was open to people on Pro, on macOS.

Worth knowing if you have a Pro subscription and had assumed the tab was not for you.

Two older models, Claude Opus 4 and Claude Opus 4.1, were taken out of the model selector — the dropdown where you choose which Claude you are talking to. They simply stopped being on the list.

This is what a model retirement actually looks like from the reader’s side. Nothing breaks and nothing needs fixing; a name you may have recognised is no longer offered, and you pick from the current ones instead.

Cowork turned up for the first time here, as a research preview — an early, deliberately limited release — inside Claude Desktop on macOS. Only Max subscribers could open it, and only on a Mac.

This is where Cowork begins. Everything later in this log about Cowork on other plans, on Windows, or on the web traces back to this date.

Claude in Chrome — Claude working inside the browser rather than in its own window — became available to everyone on a paid plan, still labelled beta.

Beta here means released but expected to have rough edges. It is a separate thing from the Desktop app, and using one does not require the other.

Claude for Excel — Claude working inside a spreadsheet rather than in a separate window — went into beta, meaning it was released but still early. Pivot tables were included in what it could handle.

For analysts and finance work, this is the version of Claude that meets you where the numbers already are. Beta is worth taking at face value, so check its output the way you would check a colleague’s.

Every conversation has a limit on how much Claude can hold in mind at once. Before this change, a long chat would reach that limit and the earliest parts would simply drop away — which is why a session you had been working in all morning could suddenly seem to forget what you agreed at the start. It was baffling, and it was not you doing anything wrong.

Context compaction changed the behaviour. Instead of dropping older turns, Claude compacts them: it condenses what came earlier into a shorter form and keeps working from that. The conversation carries on rather than hitting a wall. Detail from the oldest part of a very long chat still gets thinner over time, so it is worth restating anything that really matters — but the abrupt amnesia is gone.

Claude Opus 4.5 was released. For anyone using the app rather than building on the API, a model launch means the engine underneath got newer, and there is nothing to install or switch on.

Its full model id is claude-opus-4-5-20251101, which is the sort of string you only ever need if you are calling Claude from code.

Memory became available to Max and Pro subscribers. It is the feature that lets Claude keep hold of things about you and your work between conversations, rather than starting from nothing each time.

At this point it was limited to those two paid plans. Later entries in this log widen that.

Haiku is Anthropic’s fast, light line, and version 4.5 landed here. It holds up to 200,000 tokens of context — roughly, how much text it can keep in mind at once — and costs $1 per million tokens in and $5 out, which only matters if you pay per use through the API rather than by subscription.

This one aged well: it was still a current-generation model in August 2026, long after several newer Opus and Sonnet releases had come and gone.

Up to this point Claude replied in the conversation and you copied the result somewhere useful. From here, everyone on a paid plan could have it create and edit files directly.

It is a small-sounding line with a large consequence: it is the difference between an assistant that describes a document and one that hands you the document.

Sonnet 4.5 was released. As with any model launch, the practical meaning for someone using the app is that the engine underneath improved and no action was needed.

claude-sonnet-4-5-20250929 is its formal id, used when Claude is called from code.