A week in Generative AI: Skills, Veo & Haiku
News for the week ending 19th October 2025
Weâve had a couple of really big weeks in AI recently, so this week feels a bit quieter, but there was some interesting and significant releases nonetheless. The week mostly belonged to Anthropic with two major announcements: Agent Skills and Claude Haiku 4.5. Agent Skills is Anthropicâs solution to one of the biggest challenges in AI agents: giving them the specialist knowledge they need to perform specific organisational tasks effectively. Meanwhile, Claude Haiku 4.5 offers similar performance to Sonnet 4 at a third of the cost and twice the speed. Thatâs impressive progress in just 5 months!
In Web 4.0 news, shopping at Walmart will soon be available in ChatGPT, AI is impacting Wikipediaâs traffic, and Perplexity has paused new advertising deals to reassess its ambitions.
On the Ethics front, California became the first state to regulate AI companion chatbots, there are concerns that a recent ChatGPT update is giving more harmful answers, and OpenAI paused Sora generations of Martin Luther King Jr.
In Long Reads Simon Willison has a great piece arguing that Claude Skills might be a bigger deal than MCP, which is definitely worth a read.
Anthropic introduces Agent Skills
One of the biggest challenges in developing AI agents is bridging the gap between what an AI model can do out of the box versus the specialist skills required for specific tasks and jobs within an organisation. Itâs always been clear to me that we shouldnât expect AI models, despite being generally very capable, to excel at specialised organisational tasks without some scaffolding and engineering effort to make them work effectively.
Agent Skills is Anthropicâs answer to this challenge: a user-friendly way to package the specialist knowledge Claude needs to perform specific tasks well. Skills allow users to package step-by-step instructions and reference documents together so that when you ask Claude to perform a task it automatically loads and applies the appropriate Skill and has access to all the right information to perform the task.
You might not know it, but Skills were already being used by Claude to create files like spreadsheets and presentations. Now users can build their own Skills for the specific tasks theyâd like Claude to complete for them. Weâve seen lots of new features from the large frontier AI companies over the last few months focused on helping users build their own agents, which I think will be very useful for getting AI agents to meaningfully contribute in a work environment.
However, I also think there needs to be a much bigger engineering effort in most organisations to build the scaffolding on top of which AI agents will work to ensure they have access to all the right data and systems while maintaining appropriate controls over what they can and cannot do.
Google DeepMind releases Veo 3.1
Google DeepMind released the next version of its impressive Veo 3 video model this week with improved audio, more granular editing controls and more realistic video outputs.
Where these videos models are going is starting to come into focus - itâs not just about the realistic quality of the outputs, but the ability to control and edit the outputs. Iâm sure these models and their outputs will look incredibly quaint in a couple of years, but thereâs a clear development path to get to something truly amazing.
Anthropic Introduces Claude Haiku 4.5
Following the launch of Claude Sonnet 4.5 a few weeks ago, Anthropic has now released their smaller version of their 4.5 model, Claude Haiku 4.5. According to Anthropic it offers similar performance to Sonnet 4 at âone third the cost and more than twice the speedâ.
Claude Sonnet 4 was only released in May this year, so in 5 months Anthropic have been able to significantly reduce the costs and double the speed of the same level of intelligence. This is much faster than Mooreâs Law!
Anthropicâs CPO Mike Krieger has positioned the new Claude 4.5 family of models as being perfect for opening up new categories of whatâs possible with AI in production environments with the larger Sonnet model handing complex planing while Haiku powers smaller sub-agents that perform simple tasks quickly. This will quickly become a common development pattern as the world builds out more capable AI agents.
Web 4.0
ChatGPTâs mobile app is seeing slowing download growth and daily use, analysis shows
Perplexity Pauses New Advertising Deals to Reassess Ambitions
AI Ethics News
California becomes first state to regulate AI companion chatbots
The real reason Google DeepMind is working with a fusion energy startup
ChatGPT âupgradeâ giving more harmful answers than previously, tests find
Olivia Williams says actors need ânudity riderâ-type controls for AI body scans
OpenAI pauses Sora video generations of Martin Luther King Jr.
Long Reads
Simon Willison - Claude Skills are awesome, maybe a bigger deal than MCP
Artificial Ignorance - Revisiting 'âIntelligence Driftâ
âThe future is already here, itâs just not evenly distributed.â
William Gibson
This article comes at the perfect time. The progress with Agent Skills and Haiku 4.5 is just mind-boggling. I wonder what's next; it's truly fascinating to see this evovling so fast.