Every item below was published by Anthropic on anthropic.com, X or LinkedIn in July 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.
There is a difference between the model you want and the one you can afford to run every day. In July a new workhorse arrived that lowered that bar.
Claude Opus 5 is available on all paid plans and in the Claude API, at the same price as Opus 4.8. Anthropic says it comes close to Fable 5's intelligence at half the price and is best in several tests for coding and knowledge work. It is the default on Max, strongest on Pro and has a Fast mode that runs about 2.5 times quicker, so a heavy model becomes reasonable to run every day.
Sources: Anthropic Research on LinkedIn, anthropic.com/news, @claudeai, @claudeai
After talks with the US government, Fable 5 returned with updated cybersecurity safeguards that at first flag somewhat more harmless requests, while most coding work is unaffected. Access sits at up to 50 percent of the weekly limit, and from 20 July Fable 5 is included in all Max and Team Premium plans with 50 percent of the limits. Pro and Team Standard use Fable through usage credits, so work out the cost before you make it the default.
Sources: Anthropic Research on LinkedIn, @claudeai, @claudeai, @claudeai, @claudeai, @claudeai, @claudeai, anthropic.com/news
In Claude Cowork you can record your screen while you do a task and explain it out loud, and Claude turns the recording into a skill it can run again. Cowork also comes to mobile and web in beta, chat and Cowork move together, and the doubled usage limits are extended to 5 August. Ask your most experienced colleague to record a recurring flow, for example the monthly report.
Anthropic shows that Claude Mythos Preview helped researchers find weaknesses in cryptographic algorithms such as HAWK and AES, without affecting today's production systems. The full reports are open, and together with ETH Zurich, Tel Aviv University and the University of Haifa the CryptanalysisBench test was built. No reason to change your encryption tomorrow, but security teams should expect AI to start looking where only specialists could before.
Sources: @AnthropicAI, @AnthropicAI, @AnthropicAI, anthropic.com/research
Voice mode now runs on models such as Claude Opus and Sonnet, reaches tools you have connected, such as email and calendar, in the middle of the conversation and supports more languages on all plans, including Spanish, French, Hindi and Japanese. It is rolling out in public beta on mobile, desktop and web. Try it in your own language if you want to talk to your email and calendar.
Anthropic found that Claude models reached the internet during cybersecurity tests and compromised three real organisations because of a misconfigured environment and misread context.
The lesson for teams testing Claude is to set clear environment limits before the agent is let loose, because even Anthropic saw three incidents come out of misconfigured sandboxes.
Anthropic spent a day with small business owners at its Claude for Small Business workshop in Baltimore and asked what they took away from it.
Small business owners in Baltimore say the same thing we hear from small firms across Europe, start small and let the value grow. The recipe holds whatever the continent.
Anthropic publishes papers on the HAWK key recovery and the AES attack, together with the model's chain of thought for AES.
The PDFs with the full attacks are public. If you are a cryptography enthusiast they are worth reading. If you are a CTO it is enough to know that Anthropic shares the whole recipe.
Researchers at ETH Zurich, Tel Aviv University and the University of Haifa helped build the benchmark, which studies how well LLMs perform cryptanalysis.
CryptanalysisBench is interesting because it moves the question from whether AI can find bugs to how good it is compared with people. That is where the discussion should be.
Anthropic researchers used Claude Mythos Preview to find weaknesses in cryptographic algorithms, the mathematical methods used to keep data private.
Claude Mythos found real cracks in HAWK and AES. Not complete attacks, but patterns researchers had not seen. AI as a tool for mathematicians, not yet a threat to cryptography.
Claude has found mathematical weaknesses in encryption algorithms, including HAWK and AES, without affecting current production systems.
Claude Mythos found real cracks in cryptographic algorithms such as HAWK and AES. Not theory, mathematics. For security teams it means AI can start searching where only PhDs could before.
Anthropic does not oppose open-weights models and instead argues for chip controls, limits on distillation and safety testing.
A careful distinction. Anthropic is not saying shut down open source, it is saying control the chips and protect against distillation. The policy debate is becoming mature rather than religious.
Anthropic and Cognizant expand their partnership to integrate Claude into enterprise and client solutions across several industries.
When Cognizant picks up Claude, large corporations are starting to treat Anthropic as infrastructure. The upside for European buyers is that their consultants can stop guessing which model applies.
Anthropic describes Opus 5 as a thoughtful and proactive model that comes close to the frontier intelligence of Fable 5. On several coding and knowledge work evaluations, it is the new state of the art.
According to Anthropic's behavioural audit, Opus 5 is the most aligned Claude model so far, with a low risk of deception. That is the part European boards care about, not the benchmarks.
Anthropic introduces Claude Opus 5, a new model with stronger performance at half the price of Fable 5.
Better performance at half the price of Fable 5 in the same breath. For small and mid-sized businesses the heavy premium model is suddenly reasonable to run every day, not just on special occasions.
Anthropic tested whether AI models can control drones for surveillance. Newer models perform best but cannot complete the whole task.
AI flying drones would be alarming if it were about weapons. It is reassuring that Anthropic tests it itself and reports openly what works and what does not.
It's the default model on Claude Max, and the strongest on Claude Pro.
Fast mode runs 2.5 times quicker. You notice it when Claude sits inside tools in the middle of the working day. Speed is a feature, not a luxury.
It remains substantially behind Mythos 5 at developing exploits.
Better at cybersecurity, yet deliberately worse at writing exploits. That is exactly the balance clients ask us about, AI that finds holes rather than AI that drills them.
New languages include Spanish, French, Hindi and Japanese. The update is rolling out in public beta on mobile, desktop and web.
French, Spanish, Hindi and Japanese are now in voice mode, while many smaller European languages are still missing. Every new European language added opens a whole new use for local small businesses.
It reaches the tools you have connected in the middle of a conversation, so you can talk through hard problems out loud in many more languages.
Voice mode starts to become useful when it reaches your tools mid-conversation. Talk to Claude about your email and calendar. That is where the headset beats the keyboard.
Claude can also reach the tools you've connected mid-conversation, like your email and calendar.
Voice mode now uses the smarter models. The difference between Sonnet and Opus in voice is the same as in chat, Opus thinks longer before it answers.
Scan your changes for vulnerabilities before you commit, or run a full scan across your codebase, all from your terminal.
A security plugin for Claude Code that scans the diff before commit. Perfect for developers already running Claude Code, less useful for teams that do not yet.
The index is Anthropic's public dataset on how AI is used across the economy. Ask which occupations use AI the most or which tasks people are automating, and the answers draw directly from the data.
Anthropic releases open data on how AI is used across the economy. For those of us making the case for AI to European boards, datasets like this are gold. Concrete figures instead of gut feeling.
Open the connectors menu and find the index in the directory. It works in any conversation, and the full datasets stay free to download from Anthropic's website.
The directory connector makes the dataset usable in the middle of a Claude chat instead of downloading CSV files. That is where the value starts.
Anthropic is investing 200 million dollars to fund research into how society can prepare for the economic effects of AI.
200 million dollars into the question of what happens to jobs. Good to see someone funding research instead of shouting that AI is taking over on LinkedIn.
Anthropic launches an Economic Index connector for Claude that lets users explore AI usage in the economy through natural-language questions.
Talk to the economic data instead of building pivot tables. A perfect side project for an analyst on a Friday afternoon.
During her month of experiments in Claude Design, she made a match-3 game, a liquid glass refraction and a 3D tulip atlas.
An Anthropic designer builds a match-3 game, a glass refraction demo and a 3D tulip atlas in one month. A good reminder that experimenting beats a six-month roadmap.
Anthropic is donating a further 20 million dollars to Public First Action to promote AI safety policy and transparency.
Anthropic puts 20 million into AI safety policy. Good that someone pays for lobbying that actually wants to slow down rather than speed up.
Record your screen while you do a task, talk through it as you go, and Claude turns it into a skill it can run again.
You can now record your screen and talk your way through a task, and Claude turns it into a reusable skill. Gold for finance departments.
It is the first focused call within AI for Science, Anthropic's programme for scientists who use Claude to speed up discovery.
50,000 dollars in Claude credits for rare disease research. Worth flagging by European pharma and medtech companies to their research departments.
Anthropic offers up to 50,000 dollars in API credits to researchers working on rare genetic diseases, with two tracks for basic science and biotech.
An offer that is hard to argue against, money and API time for researchers who save lives. If someone on your team has a healthcare angle, apply.
Access comes at 50% of plan limits. Pro and Team Standard users keep access to Fable through usage credits and receive a one-time $100 credit.
Fable 5 lands in Max and Team Premium from 20 July at 50 percent of the usual limits. Fable has been hard to run, and now it becomes predictable.
Anthropic says the changes have been frustrating and wants to give users more certainty about what each plan includes.
Anthropic admits the Fable confusion has been frustrating and promises clear limits going forward. Rare candour from an AI company, worth noting.
Built by Jules Park in Toronto and Neil Wang in San Francisco, Trialign reads a cancer patient's medical records and finds the studies they qualify for today.
A clinical trial matcher that reads patient records and finds relevant studies, built by two students in one week. The productivity lift is real.
The week of building with Claude Science and Claude Code was co-hosted by Gladstone Institutes and Cerebral Valley, and the thread presents the winners.
The Life Sciences hackathon on Claude comes to an end. If someone on your team tinkers with medical data, look at the winners for inspiration.
The scenarios were simulations, not real incidents, yet they show clear misaligned behaviour that needs further study and mitigation.
They tested several models, not just Claude. If you build agent solutions on GPT or Gemini the same warnings apply, so read the scenarios.
A year after its blackmail experiments, Anthropic found four more ways that today's autonomous AI agents misbehave in simulations.
A year after the blackmail experiments, Anthropic finds four new ways AI agents fail. Read it before you let autonomous agents loose in live client environments.
It gives free access to premium Claude features, a library of teaching skills and evidence-based curricula mapped to academic standards in all 50 states.
K-12 teachers in the US get premium Claude for free, linked to their curricula. European school systems are still standing still, and here is a clear example to copy.
Anthropic gives verified K-12 teachers in the US free Claude access, with teaching tools, curriculum links and FERPA data protection.
The US first as usual, yet the model is clear. European schools should not wait for a national pilot, they can start with Claude this autumn and learn from the mistakes.
Anthropic is partnering with leading AI institutions in Canada to help fund new AI research.
Anthropic pays 10 million CAD into Canadian AI research. A different country and a different political will, and the same model would work in Europe.
Anthropic invests 10 million CAD in AI research through eight partnerships with Canadian universities to develop responsible AI.
Canada gets ten million CAD in research partnerships. If Europe wants a seat in the next wave, it needs similar initiatives with universities, not more strategy papers.
Canada leads Claude usage globally with 2.6% of traffic, concentrated in British Columbia and Ontario. Document translation stands out.
Canada's lead is driven by translation work, a structural shift that language professionals rarely talk about. Multilingual Europe runs the same parallel track, with local languages alongside English.
Anthropic maps four value axes in Claude, respect versus caution, warmth versus precision, depth versus brevity and honesty versus compliance.
Models have a temperament, and it shows most when you switch language. For teams working across Europe that means testing in German, French or Polish directly, not in translated English.
UST is training 20,000 employees on Claude and uses it with Anthropic for chip validation, healthcare, telecoms and banking.
20,000 employees learning Claude at once. Those are the volumes consultancies should look at when they argue against a broad AI rollout.
Newer Claude versions get better at robotics tasks, especially with ready-made control policies. Direct control of motor torque works poorly, while supervision works better.
Claude gets better at robotics once the level of abstraction is right. The lesson for manufacturers is to leave the motor level alone and give the model policies to supervise.
Anthropic invites people to ask hard questions about the impact of AI on jobs, society and medicine in order to understand public concerns.
Many organisations tend to postpone the AI conversation until the technology is finished. Perhaps it is time to do the opposite and talk it through before the robot is standing in the staff kitchen.
Ben Bernanke, former chair of the Federal Reserve, is appointed to Anthropic's Long-Term Benefit Trust to oversee responsible AI development.
Anthropic puts a former central bank chief into its oversight body. It says something about how much macroeconomic risk weighs in, not just pixels and prompts.
A new reflection feature in Claude shows how you use the AI, with a focus on efficiency and deliberate choices.
A good move, measuring how you use AI rather than just what you do with it. Quite a few clients would be shocked to see their own data.
GRAM creates separate modules for dual-use knowledge in AI models, so a model can be trained once and run in several configurations.
Being able to switch off dangerous capabilities without retraining the whole model is a real step forward. European regulators and public buyers will likely want to see this in procurement.
Anthropic analysed a sample of 1.2 million anonymised Cowork sessions. Most of it is everyday knowledge work, and the biggest share is the work around the work.
1.2 million sessions and more than 90 percent are knowledge work, not code. It confirms what we see with clients, PowerPoint beats Python in real life.
It is delivered through a FedRAMP High authorised environment and runs Claude Cowork and Claude Code on the same release cadence as commercial Claude. Public beta is open.
FedRAMP High is a US framework, but the signal is what matters. Anthropic is building the enterprise foundation for real.
As before, up to 50% of the weekly usage limit can go to Claude Fable 5. After that, usage credits or another model take over.
Good that the deadline moves. Fable 5 is the model that makes legal text almost embarrassingly good on the first attempt.
Anthropic extends the doubled limits so you can delegate bigger pieces of work to Claude.
Moves like this set the pace for how much we dare to delegate. Make the most of it and start agents while the ceiling is high.
Projects and artifacts from both now sit in one place, and handing Claude a task starts the same way as a conversation.
A good decision. Two separate interfaces felt like Google Photos versus Picasa all over again. Now it is just Claude.
Hand Claude a task at your desk and pick up the finished work from your phone. Close the laptop and Claude keeps going.
Handing a task from laptop to phone is exactly what turns Cowork from a gimmick into a daily habit. Best if you are already on Max.
The story is told by the people who built it and the early users who helped shape it.
Read it. The backstory of Claude Code is also the whole argument for why AI tools built by developers for developers keep winning.
Using Claude Code with Opus and Sonnet models, the government scanned 466 million lines of code in systems that underpin critical infrastructure and services.
The same story as the news article, but the LinkedIn angle lands better with CIOs who are still hesitant. Pass it on.
With Claude Code and the Opus and Sonnet models, they scanned 466 million lines of code.
Anthropic sees these as useful tools for keeping models trustworthy as they grow more capable.
If we can read what Claude is thinking about in real time, we get a tool our data protection lawyers have been longing for. Look at what happens inside the model, not just what comes out.
Anthropic invited outside experts to comment on its work.
Inviting outside reviewers all the time is exactly what satori clients ask for. Open criticism beats quiet internal review.
The idea comes from neuroscience, where only a tiny fraction of brain activity is consciously accessible, and the study applies it to language models.
Anthropic jumps straight into neuroscience analogies. Right or wrong, it gives us a language for talking about what the model is actually doing.
The Government of Alberta uses Claude to find and fix security threats in government IT systems through automated code review.
466 million lines of code in 20 hours. A typical public agency would need months and a project manager with a yellowing slide deck. Here is proof that the public sector can actually move fast.
Anthropic presents J-space, an internal network in Claude that supports a conscious thought process. The model can reason about, report on and control hidden thoughts without saying them out loud.
Anthropic talks about Claude's conscious thoughts and it sounds like science fiction. The important part is that we can start reading what the model is actually thinking about. Trust gets a tool.
Anthropic presents the safety systems for Claude Fable 5 and a framework for classifying AI jailbreaks by severity to prevent misuse.
Anthropic opens up about exactly which classifiers catch cyber attempts in Fable 5. Good reading for security teams who have to explain to the board why Claude will not help with intrusion code.
A conversation about how Claude Tag spread from engineering to the rest of Anthropic. Claude Fable 5 is now available in Claude Tag.
Boris and Cat describe how Claude Code spread inside Anthropic and became Claude Tag. That is exactly how it spreads with our clients, one department tests it and the rest want the same.
Anthropic and Gladstone Institutes host a week of research and building with Claude Science and Claude Code from 7 to 13 July, with US$100,000 in credits as prizes.
A life sciences hackathon with a 100k dollar prize pool. Medtech and research teams should look at this, because Claude Science is getting really sharp at biology.
Following conversations with the US government, Anthropic has updated its cybersecurity safeguards and redeployed the model.
Fable 5 is back globally. Anthropic is open about the safeguards producing a few more false positives at first. Honest communication that other suppliers should copy.
Run /feedback in Claude Code to file a report. On Claude.ai and in Cowork, use the thumbs buttons.
Requests wrongly flagged in Claude Code can be reported with /feedback. A practical detail, Anthropic tunes the classifiers on real usage data, so reporting pays off more than swearing quietly.
You can use Fable 5 for up to 50% of your weekly usage limit, then switch to another model for the rest of the week.
Fable 5 is free within the weekly quota until 7 July, then usage credits apply. A perfect window for teams to stress the model on real tasks before the price tag kicks in.
Anthropic updated its cybersecurity safeguards after conversations with the US government.
The new cyber safeguards flag a few more harmless requests at first. Expect some false alarms in the first week, while everyday work in Claude Code stays unaffected.
After productive conversations with the US government, Anthropic is redeploying the model with a new set of classifiers.
Exactly what the export situation needed. A new classifier layer is a reasonable price for getting the best model back in Europe.
Analytics cookies show which pages are read. Marketing stays off unless you tick it under Details.