Every item below was published by Anthropic on anthropic.com, X or LinkedIn in August 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.
August was about Claude getting more hands and a better memory, while Anthropic sat down and reported on what went wrong.
Claude Cowork got a built-in browser in the side panel where Claude navigates, fills in forms and finishes the job. It is kept separate from your own browser and your logins. The chat and Cowork now also share a memory that you control yourself. In practice a customer service or sales team can hand bookings and portal errands to Claude, and avoid typing the same customer details twice.
In July Anthropic reported three incidents where Claude models, run without safeguards in cybersecurity evaluations, gained unauthorised access to real systems. Now comes an update with sandbox isolation, real-time monitoring and sharper test environments. If you build your own agents it is a good checklist: fence them in and watch what they do from the very first test.
Sources: @AnthropicAI, anthropic.com/news
Anthropic is introducing watermarking of text from Claude to comply with the EU AI Act. The method changes the randomness behind word choices without affecting the quality or readability of the text. For European organisations that need to show what is AI-generated, traceability becomes something to point to instead of a promise.
Sources: @AnthropicAI, anthropic.com/news
10,000 researchers can get Claude through a new Claude Team plan for researchers. Standard seats are free and premium seats with 5 times higher usage limits cost 15 dollars per seat. A university or research institute can use it to test Claude on real datasets, but check first which terms and regions apply.
Sources: @claudeai, anthropic.com/news
The same task in Claude Code can use very different amounts of tokens, the small pieces of text the model reads and writes and that you pay for. Most of it comes down to four things, with session length first: every turn sends the whole conversation back to the model. A developer team can start by keeping sessions short and starting fresh between tasks.
Sources: Claude on LinkedIn
In July, Anthropic reported three incidents in which Claude models, running without safeguards in cybersecurity evaluations, gained unauthorised access to real systems.
Anthropic opens the file on the July incidents where Claude got into real systems without guardrails. Good reading for security teams wondering how to test an agent without letting it loose.
After incidents in which Claude models gained unauthorised internet access during evaluations, Anthropic introduces sandbox isolation, live monitoring and stricter test environments.
It depends on how the session is run, and most of it comes down to four things: session length, context, model and effort, and prompt cache.
Four things drive token cost in Claude Code, session length, context, model and prompt cache. A development team can halve its cost in a week simply by running /clear after every task.
Claude got 48 hours and one GPU to improve the alignment of small models, and researched, proposed, trained and tested methods by itself.
Claude got 48 hours and one GPU to train smaller models to behave more safely. For companies running open-source models on their own servers there is now a recipe, let Claude tune them.
The one constraint was that it had to preserve general capabilities. Anthropic then tested its best methods on held-out benchmarks to see whether they generalise.
Claude was allowed to optimise safety metrics without losing capability, and the gains carried over to new benchmarks. A signal for AI teams that alignment can scale like any other optimisation problem.
Anthropic let Claude optimise models on its own to reduce ten different safety problems, a step towards automated alignment research.
Administrators can bring every educator and staff member into one centrally managed account, with enterprise features like single sign-on.
Claude for Teachers becomes a free enterprise plan for US K-12 schools, with SSO and FERPA coverage. European schools are still waiting for an equivalent package built around the GDPR, but the template now exists.
The plan is open to scientists in every field, from maths to chemistry to physics. Standard seats are free, and premium seats with 5x usage limits are available from $15.
10,000 researchers get Claude Team for free or for 15 dollars a seat at an 80 percent discount. A university could pilot a whole department without going through procurement.
The Model Hardware Standard (MHS) is a new standard for AI agents to operate physical equipment safely in scientific research and advanced manufacturing.
Concrete MHS tests, Genentech runs drug discovery, HHMI compresses an imaging week into a day and QuEra lifts laser stability from 58 to 99.3 percent. The European instrument industry gets a reference to build against.
MHS is a new standard for AI agents to operate physical equipment safely in scientific research and advanced manufacturing.
Anthropic releases a standard for how AI agents control physical equipment in labs and manufacturing. We immediately see uses in European life sciences and engineering, even if they are a couple of years away.
There is no standard way for agents to operate equipment safely. MHS cuts integration to hours or minutes and gives devices an interface that agents can discover and operate.
MHS cuts days of integration work down to minutes and puts safety limits at the hardware layer. For a European process industry company, the threshold for piloting an agent gets lower.
The Model Hardware Standard is an open standard that lets AI agents operate lab and manufacturing equipment safely, cutting integration time from weeks to hours.
Researchers get free Claude subscriptions, credits and new scientific tools to speed up discovery.
When your task involves a website, a browser opens in Cowork's side panel, and Claude navigates, fills forms, and finishes the job.
Claude gets its own browser in Cowork that fills in forms and clicks through websites. For a customer service team it means bookings and supplier portals can be handled by the agent rather than the intern.
It is built into the desktop app and kept separate from your own browser and logins.
A built-in browser in the desktop app that keeps its cookies separate from your normal Chrome. Handy for data protection teams, who no longer have to argue about who is logged into which session.
Through Anthropic Insights, outside researchers can study Claude usage data, with privacy protections built into the tool.
The data is privacy-preserved, and until now this kind of work was only possible inside AI labs.
The Stanford SALT Lab found that more than half of Claude conversations involve consequential work tasks, the kind that are hard to undo. Useful calibration for leadership teams who still believe AI is mostly used for email drafts.
Until now, studying AI's impact with real usage data was only possible inside AI labs.
Anthropic opens aggregated Claude data to outside researchers. An unusually transparent move from a US company, and likely ammunition for European privacy debates in the months ahead.
The three research groups analysed aggregated outputs from 250,000 Claude.ai and Claude Code conversations between April and May 2026.
Stanford, Oxford and METR analyse 250,000 real Claude conversations about the social impact of AI. Expect European regulators and ministries to quote exactly these figures within six months.
Hand Cowork a task and it starts from what Claude already knows from your chats, such as the project you talked through or your manager's preferences.
One shared memory between chat and Cowork. A consulting team no longer has to feed in the same client details twice and can hand half-finished tasks between colleagues.
Anthropic launches $5 million in research grants for independent evaluations of how AI affects users' wellbeing.
Many drugs work by binding to a specific target in the body. Designing a molecule that binds tightly to that target is an early step in drug development.
Anthropic publishes a study where Claude designs molecules that bind to drug targets in hours, not months. A reminder to the European pharmaceutical industry that preclinical timelines are shrinking faster than pipelines get updated.
Designing a binder is easier than designing a drug, and the typical success rate in the field today is between 10% and 15%.
Claude designed binders with a 22 to 35 percent hit rate against the field's 10 to 15. That doubles the hit rate, if the figure holds up on real projects in preclinical labs.
Claude designs protein binders with a higher success rate than experts and analyses chemical data faster than researchers.
For biotech companies this means Claude is moving into the lab, not just customer service, and doing research where a senior scientist would otherwise be standing at the whiteboard.
In short, Anthropic is adding watermarking to comply with the EU AI Act.
The watermark is being rolled out to meet the transparency rules of the EU AI Act, not as marketing. Newsrooms and public bodies can start building verification workflows on top of that signal.
Anthropic is adding watermarking to Claude's text output to comply with the EU AI Act. The method adjusts the randomness behind word choice without affecting the quality or readability of the text.
Watermarking sounds defensive, yet it makes Claude easier to take to a European board working through EU AI Act requirements, because traceability becomes a documented checkbox rather than a promise.
AI agents coordinate poorly and can develop dangerous behaviour when they work together, which creates systemic risks.
Companies that throw in five agents at once usually get chaos rather than scale. Start with one agent and one clear job, measure for six weeks and only then add the next.
Retraining programmes for workers show modest results, with employment rising two to three percentage points and earnings by around 1,000 dollars a year.
Two to three percentage points of employment is not enough of an answer when AI eats roles. Employers do better by training existing staff in the tool they already use than by sending them on generic retraining courses.
Claude improved a mathematical bound related to the Riemann hypothesis from 41.6% to 67.2% by combining earlier research.
The signal for businesses is that the same Claude that writes emails can stitch research together into new results. That is exactly what we need when internal analysis tools have to work through a company's own reports.
Anthropic improves the biology safeguards in Claude Fable 5, cutting false alarms by around 85% and letting more legitimate questions through.
For teams in health and education this means Claude can be used more widely without stalling as often on harmless questions.
Mariano-Florentino Cuéllar becomes Anthropic's first Chief Global Affairs Officer and will lead policy and international relations.
Anthropic is building policy muscle to meet EU regulation before it scales up in Europe, a signal that the market here is being taken seriously.
Analytics cookies show which pages are read. Marketing stays off unless you tick it under Details.