Anthropic: scientists use Claude most where AI is strongest, and bench science least
Anthropic analysed 207,049 Claude and Claude Code research sessions (2 August–30 September 2026) and published the results on 2026-11-05. Data analysis is the largest use (33.7% of Claude sessions, 48.1% of Claude Code sessions); use is highest in computer science, maths and statistics and lowest in health and lab-based physical sciences, and tasks AI handles best get two to four times the use that researcher time alone would predict. Anthropic's scenarios put the speed-up from current capabilities at about 1.35x–2x, and at roughly 15x for fully automated non-physical work in computer science and statistics versus only 3x–4x in bench sciences.
Sources & depth
Google: stale TODO comments send coding agents on costly detours
Google's developer blog (2026-11-06) describes "agent distraction": in about 13% of coding-agent sessions the agent read an inline TODO(bug/...) reference, and 6% of those exposures led to follow-up work, 83% of it unprompted. Of those detours 41% chased bugs already closed and 57% pursued open bugs, widening the task; per session they added 70% more steps, 77% more tokens and 105% more waiting time. Google's fix is agents that delete only TODOs verified as obsolete, which removed thousands with over 95% reviewer acceptance.
Sources & depth
Perplexity launches "teammates", persistent agents for ongoing work
Perplexity introduced teammates on 2026-11-04: personal agents for continuing work such as marketing, sales, support and operations. They monitor work proactively, remember goals and task history across runs, and coordinate with other teammates and Computer sessions, each in its own isolated environment. Available on the web now, with iOS coming soon for Perplexity Max subscribers; Free and Pro users can upgrade in account settings. No pricing was published on the page.
Sources & depth
OpenAI introduces BioBasicsBench for wet-lab science agents
OpenAI introduced BioBasicsBench, a 100-task benchmark measuring how well AI agents interpret raw experimental data and handle everyday wet-lab science. Unfetchable — link only; the description above is OpenAI's own listing text.
Sources & depth
OpenAI partners with Hacktron to hunt vulnerabilities in its systems
OpenAI says it is working with security firm Hacktron to find vulnerabilities, test connections with third-party services, and harden its defenses using frontier AI. Unfetchable — link only; the description above is OpenAI's own listing text.
Sources & depth