AI news: cheaper models and agents that never clock out

New models from Anthropic

Claude Opus 5.5 arrived at the end of September. According to Anthropic, it matches the pricier Claude Fable 5.1 on most tasks at a fraction of the cost. Users quickly started pushing it beyond coding, into things like 3D games and product videos.

A week later came Claude Sonnet 5.5, the cheaper sibling built for everyday work. On one of the standard agentic coding benchmarks it actually edged past Opus 5.5. That's why many developers now pair the two: Opus plans the work and Sonnet carries out the individual tasks.

Both new models are already available in Macaly.

OpenAI unveils dots agents, Europe has to wait

Most people still use AI as a chat: you type a question and get an answer. The dots agents OpenAI introduced at DevDay work differently. Each one has its own computer in the cloud, connects to thousands of apps, and keeps working even with ChatGPT closed. It looks for things it could help you with on its own, and asks for approval before taking riskier steps. xAI already offers something similar under the name Grok Bot.

Regular users in Europe won't be getting dots for now. Access is limited to companies on selected business plans, and OpenAI hasn't given a reason or a date. It's a familiar pattern: Meta's Muse agent is US-only, and the new Siri still has no EU launch in sight.

DeepSeek shows how to cut AI's memory bill

Agents often spend hours on a single task. While they work, the model keeps the ongoing conversation in a memory called the cache, which is one of the biggest costs of running AI. In September, Chinese lab DeepSeek published research on shrinking it. Its latest model needs over 400 times less cache than the company's very first one, and according to DeepSeek, with no loss in performance.

Chinese companies have limited access to Nvidia's most powerful chips, which pushes them to focus on efficiency. DeepSeek also publishes its findings openly, so the whole field can build on them. Lower memory demands mean cheaper and faster AI for everyday users too.

In brief

• GPT-6.1 Sol. OpenAI released a new model for everyday work. The company says it comes close to its strongest model, GPT-6 Astra, at a fifth of the price. It's especially strong at coding and computer use, which makes it a good fit for agents that work inside apps on your behalf.
• ChatGPT for $500. OpenAI added a Pro 500 plan with the highest usage limits yet, 25 times those of Plus. It also includes Ultrafast mode, which makes Codex respond up to 8 times faster.
• Gemini 4 Argon. Google announced a new flagship model that beats the competition on a range of benchmarks. So far only selected cybersecurity partners have access, and Google hasn't said when developers or the public will get it.
• Eleven v4. ElevenLabs released a new text-to-speech model. Voices express emotion more convincingly, sadness or anger for instance, and speak more than 90 languages. The faster Turbo variant is built for voice assistants that need to respond instantly.
• Codex in the cloud. OpenAI's coding agent now runs in the cloud, so it finishes the job even with your laptop closed. It can also run regular security checks on your GitHub code and propose fixes right away.
• ChatGPT Space. OpenAI added a shared workspace where teams create notes and documents together with AI. The dots agent can join in too, turning meeting notes into a calendar event, for example.
• Apps inside ChatGPT. Developers can now add their own interfaces to ChatGPT, so users don't have to switch between apps. Connectors can also notify ChatGPT when something happens in an app and trigger an automation.

We go through all of this in detail in the latest episode of Macaly 24, including our own experience with the new models and how we're using them in Macaly.