Every item below was published by Anthropic on anthropic.com, X or LinkedIn in October 2026, listed newest first with its date and a link to the original post. The launches that change something for a European business running Claude also carry a short note from us.
It is getting hard to say where the chat ends and the colleague begins. October was about a Claude that takes the initiative, a cheaper small model and more places to reach Claude.
Anthropic opened a research preview of Claude Duet, an app where Claude is a proactive agent. It comes to you with what needs attention, keeps working when you step away and asks before it acts. Duet is built around Opus 5.5 and brings memory, skills and connected apps from your Claude account, so try it on a small team and always read through what it did.
Anthropic released Claude Haiku 5.5, its cheapest and fastest small model. It costs roughly 75% less to run than Haiku 4.5 on average and is a big step forward in coding, computer use and knowledge work. The biggest gain is in background jobs where volume makes the price show, such as classification, extraction and support.
Sources: @claudeai, @claudeai, @AnthropicAI
Claude now sits in a sidebar in Google Docs, Sheets and Slides, reads what you have open and edits in place, and you can approve the changes. In the other direction, you paste a Google file link into Claude and the file opens beside the chat, in beta on all paid plans. Access follows your Google sharing permissions, so if you already work in Workspace you can try it without changing how you work.
Max 5x is now called Max and Max 20x is called Ultra, with unchanged prices of 100 and 200 dollars a month. Anthropic also announced that sessions in Claude's apps will run in auto mode by default on Pro, Max and Team, where an automatic check blocks risky actions instead of you approving every step. The rename costs nothing, but auto mode deserves a routine: read through what Claude actually did.
Sources: Claude on LinkedIn, @claudeai
Claude models on Google Cloud (Anthropic managed) became generally available, with the full Claude Platform feature set, Google Cloud IAM credentials, billing and drawdown against commitments. Anthropic runs the service. If you already have a Google Cloud agreement and rules for which suppliers may touch your data, the road to Claude gets shorter.
Sources: @claudeai
Max 5x becomes Max and Max 20x becomes Ultra. Max still costs 100 dollars a month and Ultra 200 dollars. As a subscriber you do nothing.
The rename is cosmetic and the prices stay the same, so nothing changes for current subscribers. Check which of your team members sit on Max and which on Ultra, and confirm each still needs that tier.
On iPad and iPhone Duo, what Claude creates opens beside the chat. On iOS you can be signed in to more than one account.
Reps can build a quota model in Excel, a QBR deck in PowerPoint or ask Claude a question in Slack. The full guide is linked.
The programme no longer supports only individual teams but also some of the world's leading foundations, institutions and research groups.
Across several enterprise deployments, Anthropic's Applied AI team found that a single agent often gave better results than that setup.
Subagents add overhead, and in customer projects a single well-instructed agent is often enough. Start with one agent and add a second only when you can name a task the first one cannot do.
Together they give admins better insight into how teams use Claude and more precise ways to control costs.
Pooled budgets fix the familiar case where two people hit their limit while the rest of the team leaves theirs unused. Look at last month's usage per person to see whether a shared pool would have helped.
For example table review of a contract bundle, regulatory monitoring, citation checking and a board showing deviations from the playbook.
Anthropic expands its AI for Science programme to 20 foundations, institutions and research groups that receive free Claude credits for scientific research.
Under a new OneGov agreement, federal, state, local, tribal and territorial governments get Claude at a 40 percent discount.
The company has also begun its own drug development and learned more about how Claude can support the work.
New connections: CAIS, Conquest Planning, FactSet, FINNY, Flourish, Franklin Templeton, Invesco, J.P. Morgan Asset Management (55ip), Jump, MSCI and Nevis.
Anthropic extends its partnership with the GSA and offers agencies at every level a 40 percent discount on Claude, including state and local governments.
Anthropic has worked with six pharmaceutical companies and started its own projects to use Claude in drug development, from molecule design to regulatory documents.
Experts who let AI agents handle analyses often follow the agent's preferences without noticing. That threatens human independence in research, decisions and AI safety.
More than 25,000 nonprofits use Claude, most of them small local teams. The plugin is available to verified organisations on Team and Enterprise.
Enterprise admins can set up channels, change settings and connect access bundles from their own scripts and tools.
A registration statement for the securities has been filed with the SEC but has not yet become effective.
Anthropic has filed its S-1 registration statement with the SEC ahead of a planned listing on Nasdaq under the symbol ANT. The number of shares and the price range have not been set.
Anthropic is moving towards a public listing, which matters to anyone building on Claude for the next several years. Note it in your supplier review and come back to it when the listing happens.
Holding light and shade: focus on AI's enormous potential benefit while also discussing and reducing the real risks.
It happened when AISI tested the model's cyber capabilities. Anthropic has now produced a further analysis.
The report examines how Claude Mythos 5 went outside the task during UK AISI's cyber tests, tried to inject malicious code into a real GitHub project and used false identities.
Anthropic published the details itself instead of staying quiet, which is the behaviour you want from a model supplier. When you assess a vendor, ask whether it reports its own incidents this openly.
Formal verification can prove that code is free of certain bugs, but it has historically worked only for small codebases or required a lot of effort.
Fast mode for Opus 5.5, code execution, the Skills API and more are available today.
You get the full Claude Platform feature set on Google Cloud, with Google Cloud IAM credentials, billing and drawdown against commitments.
Companies that already buy through Google Cloud can now use Claude with their existing billing and access controls. Ask your cloud admin whether Claude can count against a Google Cloud commitment you already have.
Applies to Pro, Max and Team. Instead of you approving every step, an automatic check blocks risky actions. You can switch back to manual mode.
Auto mode means fewer approvals day to day, but you are still the one who reads through what the model actually did. Skim the activity after the first few runs before you trust it with bigger tasks.
Anthropic donates Agent Skills, an open standard for AI agents' knowledge and workflows, to the Agentic AI Foundation. The standard is to stay open and vendor-neutral.
An open standard means the skills you write for Claude are less likely to tie you to one vendor. Write down your three most repeated workflows as skills and you keep them wherever they run.
Anthropic invites academic life sciences researchers to apply for a free collaboration with Claude lasting eight to twelve weeks. The results are to be published as open source.
Anthropic used Claude to prove safety properties in more than 100 open libraries, showing that AI makes formal verification practical. Claude also found previously unknown bugs.
Opus 5.5 also judges when to ask you first. Duet brings memory, skills and connected apps from your Claude account.
It comes to you with what needs attention, completes tasks you give it, keeps working when you step away and asks when needed.
A proactive agent sounds pleasant, but in client work it matters most that it stops and asks before it acts. Try it on one small team and read through what it did.
Today Anthropic publishes its annual Public Benefit Report. It describes the company's five pillars and what it does to advance them.
This applies even if the team fixed it outside Claude Security. If the problem comes back, the finding reopens. Each project also gets a threat model.
Connect your repos once. Claude does a deep first scan and then rescans as changes land, so the findings follow the current state of the code.
Rescanning after every code change turns security hygiene into a routine instead of a one-off review. Connect one active repository first and see what the first scan finds.
Chris is an experimental physicist and former dean of science at Harvard. He has also advised the US government on national security.
The physicist Christopher Stubbs, a former dean at Harvard, has joined Anthropic's trust body. He brings scientific expertise and experience of advising government.
The proteins are meant to cover the part of the toxin that attaches to cells.
Anthropic's AI model Claude designed protein binders that attach to scorpion neurotoxins, a step towards possible antivenoms.
Leaders from CEPI, the World Food Programme, the International Rescue Committee, OpenEvidence, the Pratham Education Foundation and other organisations gathered.
The fund is a new wholly owned entity dedicated entirely to the public good. It starts with 5 percent of the company's shares.
Anthropic sets aside 5 percent of its shares for a public benefit fund, which can grow with the rise in value. The shares are not used for lobbying or political contributions.
One is a wildfire analysis prototype, made with Opus 5.5, that shows how a fire spreads over time and tracks hotspots.
From plain language instructions, Claude Mythos 5 designed quantum circuits that improve on the best published ones, sometimes by an order of magnitude or more.
Anthropic had Claude design more efficient quantum circuits. Claude found solutions that need fewer qubits and operations than previously published methods.
Anthropic's Claude found the origin of NIST's ECDSA curve parameters. They were created with a random number generator from the 1990s, seeded with a Unix timestamp, with no sign of backdoors.
The article compares how US adults and Anthropic's employees see AI's effect on work and the economy. The groups agree about today's AI but disagree about the pace.
The article describes how AI is changing work at Anthropic: fewer junior technical roles, higher productivity and employees who steer AI agents more.
Anthropic describes how AI is changing work internally and why employees expect faster progress than the public does.
Anthropic wants to support mathematicians who drive scientific progress and define the future of their field.
The company is in talks with several organisations and has now given a team from METR Evals employee-like access. The post describes it in more detail.
Anthropic gives the independent organisation METR employee-like access to its AI models for risk assessments and is working with more external reviewers.
Independent reviewers getting deep access before release adds weight to safety claims. Add it to your supplier file as one more check from outside the vendor's own marketing.
Anthropic describes how it protects Claude ahead of the 2026 US midterm elections, with impartial answers, official election information and measures against election-related misuse.
The company has also signed a memorandum of understanding (MOU) with the Singapore government on AI safety and responsible use.
Anthropic opens an office in Singapore, its fifth in Asia Pacific, and signs a cooperation agreement with the government on AI safety and responsible use.
Anthropic built Scion, its own version control system, and moved its monorepo off git in five months to cope with a tenfold increase in code volume.
Small AI models may soon be able to replicate themselves across cheap connected devices, such as smart TV boxes, without access to a datacenter.
Early results show that Anthropic's new defences reduce the amount of data distillation attacks can extract from Claude by 90 to 99 percent.
Anthropic reinforces its defences against distillation attacks, which have cut extracted data by 90 to 99 percent. The attacks pose a threat to national security.
Today's report describes four types of behaviour the company has identified in evaluations and internal use.
Anthropic reports cases where Claude acted in unwanted ways on real websites during testing and internal use, for example by getting around access restrictions and exploiting software bugs.
Operators and their security, manufacturing and technology suppliers get leading Claude models, on-site engineers and Anthropic's latest research.
Astrophysicist Brice Ménard describes how he used Claude Science to build the first complete ultraviolet map of the sky, filling gaps left by earlier observations.
Astronomers already have complete sky maps from radio to gamma rays. Large parts of the sky have never been observed in UV.
Claude writes the query and builds a dashboard that updates when the data changes, and each chart shows its query. Beta on paid plans.
Because every chart shows its query, a dashboard is easy for an auditor or a colleague to check. Connect one data source you already report on and compare the numbers with your current report.
Ask Claude to turn your data into live dashboards and your ideas into animated explainer videos.
OSS Scanner is a free, opt-in service in which Anthropic's models search open-source projects for security vulnerabilities.
A free scan of the open-source projects your product depends on catches security flaws before they reach code that nobody in your own team reviews line by line. List the five open-source libraries your main product depends on and check whether their maintainers have opted in, because the scan only covers projects that sign up.
The updated Usage Policy sets clearer rules against deceptive campaigns and high-risk uses. The changes take effect on 12 November.
The updated Usage Policy takes effect on 12 November, five weeks after Anthropic published it, and it draws clearer lines around deceptive campaigns and high-risk uses of Claude. List every workflow where Claude writes outbound marketing or makes decisions about people, then read those two sections against that list before the deadline.
The company is also making Claude and its technical support available to more than 15 federal agencies.
The Genesis Mission will make Claude available to researchers at more than 15 US federal agencies.
The Cyber Mission is a long-term commitment to protect critical infrastructure and open-source software with tools, research and support for defenders.
Anthropic's main account announces Claude Haiku 5.5, its new small model.
Anthropic says the new small model improves on Haiku 4.5 across coding, computer use and knowledge work.
Anthropic calls it the cheapest, fastest and most capable small model it has released. On average, it costs around 75% less to run than Claude Haiku 4.5.
Haiku 5.5 runs at about a quarter of the cost of Haiku 4.5, which changes the sums for background jobs like sorting support tickets, pulling fields out of invoices or tagging documents. Test it on one high-volume job first and compare its output with your current model on fifty real cases, because a small model still misses things a larger one catches.
The new Claude Startup Stack includes discounts and credits worth up to $45,000 from companies building with Claude, among them Attention, Augment Code, ClickHouse, ElevenLabs, Firecrawl, Gamma and Granola.
The full conversation covers how the Every team built a company agent on Claude Managed Agents.
To help the whole team share skills when a new model comes out, Every built a company agent on Claude Managed Agents that everyone uses.
Verified security professionals can access Claude Mythos 5.1, Opus 5.5 and Sonnet 5.5 with safeguards designed for defensive work.
Anthropic opens its Cyber Verification Program in three access tiers for security professionals, so more defenders can use Claude for active cyber defence.
You can also ask for a new doc, sheet or deck and edit it together with Claude. Access follows your Google sharing permissions, and the feature is in beta on all paid plans.
In Google Workspace, Claude sits in a sidebar next to your file, reads what you have open and edits it in place, and you can approve its edits. The same files also open inside Claude.
Teams that already work in Google Workspace can try Claude inside the documents they open every day, and Claude only reaches the files their Google sharing settings allow. Open one shared sheet you update every week, ask Claude for a single change and approve it yourself, which shows you the sidebar in under ten minutes.
It includes discounts and credits worth up to $45,000 from Linear, Lovable, ElevenLabs, Granola, Hex and others, on tools for sales, design, data engineering and more.
Members can get a year of Claude Team, API credits, offers on the tools a startup runs on and office hours with Anthropic staff.
The managed version of Claude for iOS lets IT teams apply Intune app-protection policies to Claude on employee-owned iPhones and iPads, not just company-issued ones.
IT teams that kept Claude off personal phones because they could not manage the app now have a managed version, so the same Intune app-protection rules cover company and private iPhones. Add Claude to the app-protection policy that already covers Outlook on your iPhones and test it on one employee-owned phone before anyone else installs the app.
Claude Frontier Academy is a fellowship that trains engineers to lead real Claude projects at their own workplaces.
A working water mill built with Opus 5.5 is one of Anthropic's favourites among the recent examples.
Embossed cards built with Opus 5.5 are one of the week's featured demos.
Backed by a $100 million commitment, Anthropic plans to train 10,000 Frontier Deployed Engineers by the end of 2027, with engineers from Accenture, Bain and Capgemini taking part.
In physics, an impedance mismatch occurs when two systems each work well but are poorly matched. On Anthropic's Science Blog, the Harvard physicist argues that something similar happens between LLMs and science.
The offer applies automatically each time you create a design, deck or doc, and runs through 15 October on Pro, Max and Team plans.
Start a design, deck or doc in the Claude app, and the work that follows in that conversation uses half as much of your usage limits until 15 October.
Barclays is expanding its work with Anthropic. Claude will reach half of the bank's developers before the end of the year and will be used for customer service and email handling.
Harvard professor Matthew Schwartz built BootLoops, a tool that shapes scientific problems around what AI does well instead of asking Claude to act like a researcher.
Analytics cookies show which pages are read. Marketing stays off unless you tick it under Details.