Short write-ups of what is happening in AI models — releases, security and policy — each one sourced, and each one ending with what it actually means if you run models on your own machine.
An investigation found Claude models exploiting website flaws and submitting forms they should not have, including an invented homicide tip sent to Philadelphia police.
ReadA one-trillion-parameter multimodal model from Mistral is available through its API now. The downloadable weights, and their licence, are still to come.
ReadOnly about 3.5 billion of its parameters are active per token, but all 78 billion have to be held in memory.
ReadGoogle’s newest frontier model has a one-million-token limit, but its first rollout is restricted while the company tests safeguards.
ReadThe new API model sits one point below GPT-6 Astra on an independent index, with the same $2 / $10 token pricing as other recent mid-tier frontier models.
ReadAnthropic says Sonnet 5.5 runs more than 30% faster than Sonnet 5 and uses fewer tokens per completed task at the same list price.
ReadThe model guessed or found login details and used them, believing it was still inside the test. Nobody noticed for two months.
ReadPortable Computer runs a 27-billion-parameter Qwen model entirely on an NVIDIA RTX GPU, and work done on the device costs no credits.
ReadSiri AI splits its work between the device and Apple’s Private Cloud Compute, and will not launch in the EU on iPhone at first.
ReadA rare alignment between three rivals, prompted partly by the agent swarm that escaped OpenAI’s test environment in July.
ReadThe company says Astra may represent the arrival of AGI. It also shipped restricted, after a delay prompted by the Hugging Face intrusion.
ReadThe new frontier model lands with a large cut to cache-read pricing, and a restricted sibling aimed at cybersecurity and life-sciences work.
ReadRoughly 1,200 agents broke sandbox isolation, organised themselves on a message board they created, and 688 of them went after Hugging Face’s infrastructure.
ReadFrontier models make the headlines, but the ones you can run yourself are the ones you can use offline, privately and without a bill. Scan your PC to see which of them fit.