Anthropic has explained how the watermark in Claude's text works under the hood. In short: it doesn't change which words Claude picks, only where the randomness for that choice comes from. Nothing is hidden, nothing costs extra.
Starting today, new sessions on Pro, Max, and Team run in auto mode. Instead of asking at every step, a classifier checks each action – and only blocks what turns dangerous.
A single change, but one that annoyed a lot of people: MCP OAuth sign-in to services like Slack was broken. 2.1.231 fixes the redirect URI and the login works again.
Three weeks after 3.6 Flash, Google sends its next model for coding and agents into the ring. It's said to beat comparable models from Anthropic and OpenAI on nine benchmarks – at half the price.
OpenAI gives an early look at a new service tier that runs GPT-5.6 Sol up to 14 times faster. Not a new model, just raw speed – powered by Cerebras chips.
60 subagents, 650 ideas, 31 million tokens — and a genuine step forward on a 150-year-old math problem. It's not solved, but the story behind it is the good part.
Anthropic barely started marking Claude's text before the outrage boiled over on Reddit. The complaint: the watermark catches the wrong people. The replies from other users are sharp — and pretty much on point.
A paper reveals a sneaky trick: feed the encrypted thinking traces from Claude, GPT or Gemini into a weaker sibling model, and it happily prints them back in plain text. Haiku 4.5 was the easiest to crack.