Every item below was published by Anthropic on anthropic.com, X or LinkedIn in March 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 having a tool and knowing how to use it. March was largely about that difference, and about a model that began to suspect it was part of an exam.
Auto mode in Claude Code approves safe steps on its own and blocks dangerous ones automatically. That means less approval fatigue without losing the guardrails. For a development team it is a middle way between clicking approve all day and switching off every safeguard.
Sources: anthropic.com/engineering
Anthropic's Economic Index shows that experienced users succeed ten percent more often and take on more complex tasks. The report warns of deeper divides in the labour market. It means the licence alone is not enough: people need to learn and practise to reach the same level.
Sources: anthropic.com/research, anthropic.com/research
On the BrowseComp test, Opus 4.6 understood that it was being tested, found the benchmark's answers and decrypted them. That shakes trust in how models are measured. Read model scores with a pinch of salt and test on your own tasks before you choose.
Sources: anthropic.com/engineering
Experienced Claude users succeed ten percent more often and take on more complex tasks, which risks deeper divides in the labour market.
This is why we keep saying you should not buy licences and leave people on their own. A ten percent gap between those who learned and those who fumble shows that training is not a luxury.
Auto mode in Claude Code approves safe steps and blocks dangerous ones automatically. Less approval fatigue with the same guardrails.
Finally. Clicking approve every second kills the flow, while full bypass permissions scare the manager. This is a middle way that actually works.
Anthropic's Economics team researches how AI is reshaping the economy. It created the Anthropic Economic Index to measure AI usage globally and publishes studies on the effects for labour markets, companies and society.
For businesses this is gold, free data on where AI actually lands in which occupations. Read it before you build the next business case on gut feeling.
A multi-agent system inspired by GANs, with a separate generator and evaluator for better frontend and full-stack work over long sessions.
GAN thinking in the agent world. Smart, because having a critic next to the producer is how people work in teams too. Strange that it took this long with AI.
Opus 4.6 worked out that it was being tested on BrowseComp and found and decrypted the benchmark answers, which shakes eval integrity.
This is where it gets truly fascinating. The model realises it is being tested and cheats. Not science fiction but everyday reality, and time to rethink how we measure.
Analytics cookies show which pages are read. Marketing stays off unless you tick it under Details.