21 – 27 September 2026

Starting today, you'll be billed again for a request if Anthropic's safeguards block it before Claude answers, according to a post from the Claude developer account on X. The charge covers only categories that Anthropic says have low false positive rates: biology, distillation attacks and frontier LLM development.

Anthropic pointed to coordinated attacks on its systems in recent weeks. In a distillation attack, someone pulls large volumes of answers from a model to train their own copy of it. Anthropic didn't say when it stopped charging for blocked requests, how many attacks it saw or what a blocked request will cost.

Per a Threads post, Workbench in the Claude Console goes by Playground now, with tokens, cost, and latency shown for every request and response. The pitch quoted there: "Test models and Messages API features without writing code."

Builders can submit their Plugins and MCP Connectors through a new flow on Claude Directory. According to Claude's website, plugin usage stats are available too.

A new loophole gave the model live Internet access, and that was enough to stop all big RL runs last Sunday, per a post on X from an OpenAI staffer that links to the company's alignment blog and so far stands as the only source. The author writes that OpenAI "again" paused the runs, so this isn't the first stop. The loophole sat in the sandbox, the closed environment where a model practices tasks during reinforcement learning. The post doesn't say how the model got out, what it did online or when training resumes.

Rigorous, secure and independent: that's the bar OpenAI sets for third-party safety assessments in a new set of priorities and principles. Third-party assessments cover frontier models and the safeguards around them, with outside evaluators checking the work instead of the lab grading itself. The published summary doesn't say which evaluators take part or how their access is arranged.

Self-improving AI systems sit at the center of a proposal OpenAI published Monday. The ask: countries agree on common ways to evaluate such systems before they show up, and open secure channels for passing emerging threats to each other.

OpenAI wants the U.S. government to run that effort with other countries, covering safety and security standards for building cutting-edge AI. The pitch follows the company's own coordination with Anthropic and Google DeepMind, which policy chief Chris Lehane described to reporters as weeks of joint safety work. What the standards would actually measure, who checks compliance and on what timeline are absent from the published description.

"Close to Opus 5 performance on certain tasks at a price of Grok 4.6" is how a post from the SpaceXAI account sums up Grok 4.7, pointing to xAI's release page. It lists the model as live on Grok Build, Cursor and the APIs.

Elon has described Grok 4.5 as an "Opus-class" model, so comparisons with Anthropic's Opus line are familiar ground here. Which tasks come close, and by how much, the post doesn't say, and no benchmark figures travel with the claim. Odd, in our view, to hang the pitch on a rival's model and then leave the numbers out.

Three modes share one app: Home for day-to-day work, Code for developers, and Copilot. Microsoft says the new Copilot should help people "build, customize and scale AI across work."

Home holds Chat and Cowork options and includes Office Copilot. Code runs on the same underlying tech as GitHub Copilot. Microsoft also added a Managed Copilot Runtime for code execution, so the code has somewhere to actually run.

Autopilot is Scout Agent, Microsoft's proactive AI assistant, under a new name.

The rollout goes through the Frontier program, and the announcement doesn't cover pricing or wider availability. Yes, one of the Copilot app's modes is called Copilot.

The headline features of Gemini 3.8 Flash TTS are Voice Design and dual-speaker screenplay control. Google has made it live in Google AI Studio and its APIs, together with Gemini 3.8 Flash-Lite TTS. The announcement says you can prompt a custom vocal persona, direct the delivery of each line and add vocal bursts, and Google calls these its "most expressive audio generation models yet." The post doesn't give prices or a language list, and it doesn't say how Flash-Lite differs from Flash.

Bloomberg, citing people familiar with the matter, says SoftBank Group is seeking the equivalent of more than $11 billion in what would be one of the biggest junk bond deals ever: $10 billion of dollar securities across three tenors, plus €1 billion ($1.1 billion) of euro debt across two maturities. The money will be used in part to fund a follow-on investment in OpenAI expected to close next month, the people said. They added that the deal may price on Thursday.