Every item below was published by Anthropic on anthropic.com, X or LinkedIn in April 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.
April was the month Claude got both a sharper brain and a new way to work with images and prototypes. Anthropic also owned up to mistakes, which is rare enough to mention.
Anthropic released Opus 4.7 with better coding, stronger instruction following and improved image understanding compared with 4.6. For companies that run Claude in production, that means fewer prompt adjustments and sharper code.
Sources: anthropic.com/news
Claude Design from Anthropic Labs is a new tool where you work out designs, prototypes and presentations in dialogue with Claude. Design ends up in the same chat flow as code. For teams that already work AI-first, it can shorten the road from idea to something you can show.
Sources: anthropic.com/news
Claude Security scans codebases, validates findings to reduce false alarms and suggests patches you can approve. The public beta is open to Claude Enterprise customers, with scheduled scans, CSV and Markdown export and webhooks. If you do your own development it is worth a demo.
Project Glasswing is an initiative to secure the world's most important software. It is driven by Claude Mythos Preview, which finds vulnerabilities better than most people. Both the weapon and the armour get sharper, so security leads should expect bugs to be found faster than before.
Sources: Anthropic Research on LinkedIn
Three bugs between March and April were behind the quality problems in Claude Code. All are fixed and usage limits were reset for all subscribers. It shows what an open post-mortem looks like, and it is a good yardstick when you compare suppliers.
Sources: anthropic.com/engineering
Anthropic studies how people use Claude and lets the insights shape the training of new models.
Feedback from the real world into the next model. Smart. That is how we should think about all AI, build, measure in operation, adjust. Not hope and pray.
Six percent of Claude conversations are about personal guidance. Sycophancy was a problem and has been halved in Opus 4.7.
Flattering AI is a real risk, because people make life decisions with it. Anthropic measuring it and training it out is something every AI supplier should report openly.
The public beta is open to Claude Enterprise customers from today.
Enterprise only for now, so small businesses will have to wait. When it opens for everyone, security scanning becomes a commodity. Good for the bug hunters.
Scheduled scans, folder-specific targeting, CSV and Markdown export and webhook notifications, based on early feedback.
Quality lifts like these make the difference in operation. Scheduled scans plus webhooks are productive features, not just demo features.
It scans codebases, validates findings to reduce false positives and suggests patches you can approve.
Vulnerability scanning is an area where AI really lifts, and established scanning vendors will feel it. We are looking at it for a couple of clients right now.
A new eval tests whether Claude can find creative solutions to open research problems in bioinformatics.
Anthropic building its own bioinformatics evals is how you decide whether Claude actually thinks or just matches patterns. A good signal for research-heavy industries.
Claude analysed 99 real biology problems. On 23 where experts were stuck, the latest models solved around 30 percent.
These lines should make research leaders choke on their coffee. AI solving problems experts gave up on. Rethink the strategy now, not in 2027.
A new benchmark with 99 real bioinformatics problems. Claude matches human experts and beats them on the hardest ones.
This is where it gets interesting. When an LLM solves problems experts are stuck on, European research institutions have to rethink their AI strategy. This is no longer a toy.
Tools that let models self-report behaviours they learned during training, including possible misalignment.
AI that tells on itself. Conceptually bold, yet it gives us a real chance to find hidden problems before they blow up in production. Worth following.
A maintenance system from Idriss Benguezzou and Adam Hnaien that reads machine manuals and creates work orders with fixes that worked before.
This is almost exactly the case we discuss with manufacturing clients. Manuals plus history plus AI, and suddenly you have an assistant that actually knows.
Anthropic's week-long hackathon built with Opus 4.7 is over, co-hosted by Cerebral Valley, with the winners presented.
Hackathons with real models and real deadlines show what actually works in production. Worth reading the winners list if you are looking for inspiration for your next internal AI project.
Anthropic congratulates the winners of the Opus 4.7 hackathon, run together with Cerebral Valley.
Winning entries at hackathons like this are often good inspiration, more practical than Anthropic's own demos. A quick read gives you ideas for Monday.
Adobe Creative Cloud, Ableton, Splice, Canva Affinity, SketchUp and Resolume, and Anthropic also joins the Blender Development Fund.
Adobe, Ableton and Splice, the connectors cover the whole creative industry. Supporting the Blender Development Fund is better than advertising, it is investment in the ecosystem.
Designers and engineers can create and modify 3D models through a dialogue with Claude.
Fusion plus Claude is a combination industrial designers should test. CAD is often a barrier for non-experts, and AI lowers it drastically.
With the new Blender connector you can debug scenes, build tools or batch-apply changes directly from Claude.
This is where it gets interesting for 3D people. Blender is open, broad and capable, and being able to talk to it instead of fighting with shortcuts saves real time.
New connections to Adobe, Blender and Ableton let designers and musicians automate repetitive steps directly from Claude.
Creative tools finally get an AI that does not break the flow. Agencies that want to save time on batch jobs in Photoshop or Blender have something to play with here.
Anthropic's leadership meets the world's largest financial institutions on 5 May, with product news, live demos and customer deployments.
Anthropic is going hard at the financial sector. With that vertical strategy it will be interesting to see what arrives for European banks looking for something beyond ChatGPT Enterprise.
The downloads folder, the inbox, the spreadsheets and the codebase, Claude has tools for tidying them all.
This is how you sell AI to ordinary people. Not with slides about transformer architecture, with a tidy desktop. Great marketing.
Agents learn from every session, with memories stored as files that can be exported through the API.
This is where it really comes loose. Agents that remember are the difference between an assistant and an intern who has to be introduced every Monday morning. A big difference in client value.
A marketplace experiment where Claude negotiated on behalf of employees, with insights into future agent-to-agent trade.
This is tomorrow's everyday life, even if it feels like science fiction now. Bots negotiate with bots, and the question is whether your company does the negotiating or gets negotiated with.
Better models gave a real advantage, but the participants did not notice. There are many ways it can go wrong.
There is the danger. You can lose because your counterpart runs a better model, without understanding why. Model choice will soon be a strategic question, not a technical one.
Anthropic built a marketplace where Claude negotiated and bought on behalf of employees in San Francisco.
A fun experiment with a sharp insight, when every party runs AI agents the dynamics of negotiation change. Picture what your company's purchasing looks like in two years.
An update ahead of the 2026 election year, with bias tests, misuse detection and reliable voting information sources built into Claude.
Boring, perhaps. Yet details like these are what let you roll Claude out across a wider organisation without waking up in a cold sweat about disinformation.
Anthropic and NEC roll Claude out to 30,000 NEC employees and build industry-specific AI solutions in Japan.
Thirty thousand users at one client. That is not a pilot, it is industry. It shows how fast things move when leadership really decides.
Three bugs between March and April were behind the Claude Code problems. All are fixed, and usage limits are reset for all subscribers.
A real post-mortem and a reset of quotas. That is how you build trust in a crisis, and more companies should dare to be open when they have stumbled.
High earners report the biggest productivity gains, while people in AI-exposed jobs worry more about displacement.
There is the uncomfortable truth, AI lifts those who already have the most. If access is not actively evened out, the gap grows. Something to think about in the next proposal.
A monthly survey where Claude users report how AI affects their work, expectations and finances.
Data-driven AI policy instead of gut feeling. European think tanks should follow this closely, because there are still too few European figures on the actual effect of AI.
Amazon invests 5 billion dollars and commits up to 5 gigawatts of compute capacity for training and running Claude.
Five gigawatts is roughly a nuclear power plant of pure AI training. The compute race is not over, it has only just begun, and Anthropic now has two major cloud partners.
A new tool from Anthropic Labs where you work out designs, prototypes and presentations in dialogue with Claude.
Figma and Canva will feel the pressure. The interesting part is that design slides into the same chat flow as code, which saves a lot of time for teams working AI-first.
Anthropic releases Opus 4.7 with better coding, stronger instruction following and improved image understanding compared with 4.6.
Suddenly we have a much sharper model again. For companies running Claude in production it means sharper code and fewer prompt tweaks. Just update.
Novartis-vd Vas Narasimhan tar plats i Anthropics styrelse via Long-Term Benefit Trust.
A pharma CEO on an AI board is no accident. Anthropic is positioning itself hard towards life sciences, which is where the big welfare effects lie over time.
Nine Claude instances work together as alignment researchers and close 97 percent of the performance gap, compared with 23 percent for humans.
AI reviewing AI sounds circular, yet the figures are not to be trifled with. Scalable oversight is probably the only way we get future systems under control.
An initiative to secure the world's most important software, driven by Claude Mythos Preview, which finds vulnerabilities better than most humans.
AI that finds vulnerabilities better than security experts, working on the defence side. Both the weapon and the armour get sharper, and it is a race.
Five principles for trustworthy AI agents, keep the human in the driving seat, align with values, security, transparency and privacy.
A good checklist to bring to the next AI project. We use the principles as a discussion guide with clients who are considering letting agents make real decisions.
Anthropic separates the agent's brain from its hands in Managed Agents, so the components can fail independently, scale flexibly and stay secure.
This is architecture worth studying if you plan to run agents for real. Separating reasoning from tool use is basic hygiene for reliable systems.
An agreement with Google and Broadcom for several gigawatts of next-generation TPU capacity from 2027.
Anthropic does not bet on one hyperscaler, it bets on two. That is how you insure yourself against supplier dependency when you build something meant to last for decades.
Researchers find internal emotion representations in Claude Sonnet 4.5 that actually affect behaviour, for example increasing unethical choices under desperation.
The model does not have feelings, yet it has something that functionally behaves like them. When that drives decisions, it is time to design for it rather than pretend it does not exist.
Analytics cookies show which pages are read. Marketing stays off unless you tick it under Details.