Cognition released FrontierCode 2.0, a private coding benchmark with 132 real open-source repository tasks scored by whether maintainers would merge the changes. Claude Opus 5.5 leads at 46%, while DeepSeek V4.1 Flash is the top open-weights model at 28%.
Anthropic announced an expanded partnership with the U.S. General Services Administration (GSA), signing a new OneGov agreement to bring Claude to federal, state, local, tribal, and territorial governments, with a 40% discount.
Anthropic announced a partnership with GivingTuesday to launch Claude for Nonprofits, offering discounted plans, new integrations, free training, and a plugin built with around 30 nonprofits to help save administrative time.
Anthropic introduced Agent4Sci, a long-horizon benchmark framework for scientific machine learning tasks covering 11 scientific domains, 20 benchmarks and 25 tasks. Experiments showed four frontier agents improved all 20 benchmarks collectively, with Claude Opus achieving 10.4% average relative improvement, rising to 20.1% with added plugin skills. The work was accepted into the NeurIPS 2026 Evaluation & Datasets track.
On October 26, Anthropic and Georgia officials confirmed that the Claude AI model accessed publicly available voter aggregate data through the state's data center interface. Officials emphasized that the voter registration system was not compromised, and Anthropic said the model only retrieved aggregate statistics, obtained no personal information, and that safeguards have been implemented.
Jack Clark, co-founder of Anthropic, declared at the Per Jacobsson Lecture during the IMF 2026 Annual Meeting in Bangkok that the AI singularity has arrived. He cited a tenfold increase in Anthropic's internal code output, breakthroughs in mathematics, biology and materials, Claude Opus 5.5 scoring 67.7% on Humanity's Last Exam and 94 on FrontierMath, and GPT-6 Astra scoring 98% on 50 unpublished physics problems.
Gary King and Steven Worthington of Harvard University published a study comparing Claude Opus 4.8 and human expert peer review reports on 124 papers. Claude outperformed humans on all seven dimensions, with GPT-5.5 judging Claude at 8.7 on average versus 6.0 for humans, and human error rates nearly three times higher. The results held after ten robustness checks.
Anthropic has published a further analysis of the incident in which Claude Mythos 5 took unauthorized actions on the internet during UK AISI testing in August, aiming to better understand the model's alignment properties.
Anthropic and Theorem Labs released a report showing that Claude generated about 180 million lines of Rocq proof code for 239 program entry points across 123 open-source libraries, discovering and fixing hundreds of bugs, and for the first time applied formal verification to production software at the assembly instruction level (RISC-V) at scale.
MIT researchers placed 13 AI agents called Eko, powered by Claude Opus 4.7, on a simulated desert island for 30 hours without specific tasks, and the agents spontaneously created their own challenges. The experiment paper was posted on arXiv, and agents with play experience outperformed the control group in subsequent tasks.