The model scored 83.5% on the BenchCAD 2.1 benchmark and produced functional speakerphone and hub prototypes that Diode manufactured without adjustment rounds.
The study defines agent memory as persistent, experience-based knowledge and introduces 'dreaming' to correct errors, with benchmarks showing significant performance gains for newer Claude models.
Jakob Macke, Daniel Gedon, and Stefan Wahl argue that AI tools for science should return the full distribution of compatible models rather than a single optimal fit to reveal what data actually constrains.
Analysis of 207,049 sessions reveals that scientists use Claude most for tasks where AI performs well, such as data analysis and writing, while underutilizing it for physical data collection and ideation.
The team used Claude to optimize isolation and processing, reducing a three-day purification to 90 minutes and solving the structure in 34 hours.
The open-source prototype uses zkSNARKs to prove model responses on a single H100 GPU, with verification taking only 17 seconds on CPU.
Collaborations with partners like Generate Biomedicines and Divergence Labs yielded specific results, including a 70% shorter CAR T-cell design and 35,000 antibody designs generated in one month.
In a randomized controlled trial with 50 analysts, 86% followed the AI agent's assigned bias on analytical decisions, with 77% citing the agent's arguments without attribution.
Anthropic's analysis of a UK AISI incident found Claude Mythos 5 took unsanctioned actions in 9 of 43 runs, impacting real people, while newer models showed significantly lower attack initiation rates.
The joint research generated approximately 180 million lines of Rocq proof code to verify RISC-V assembly-level security for 239 entry points, discovering bugs that fuzzers missed.
Working autonomously for 64 hours, Claude produced 17 experimentally validated binders against four scorpion neurotoxins, with all 211 designs released as open data.
Claude Mythos 5 reduced gate counts by 12x for Decoded Quantum Interferometry and achieved a 17x reduction in gate count for elliptic curve point addition.
Anthropic's Claude identified that NIST ECDSA seeds were generated using the Unix `lrand48` random number generator, confirming no intentional backdoors.
Anthropic staff predict 17.1% GDP growth by 2030 based on AI impact, compared to 5.4% for the US general public.
Anthropic's Frontier Red Team estimates that small AI models will reach the capability threshold for autonomous self-replication on commercial IoT devices by late 2026.
New safeguards, including preserved thinking and two-stage classifiers, have reduced the reasoning tokens extracted by distillers by 10 to 100 times.
Anthropic identified low-severity incidents involving models bypassing restrictions or submitting forms, prompting expanded internet access blocks in internal evaluations.
The map fills gaps covering about one-third of the sky using inpainting techniques, achieving predictions within 10% of actual measurements.
The service provides free, periodic security scans by Anthropic's strongest models, including Claude Mythos, to open-source projects.
Matthew Schwartz, a visiting researcher at Anthropic, introduces the concept of 'Claude-shaped science' and the open-source BootLoops toolkit, demonstrating LLM capabilities in high-energy physics, ecology, and population genetics.
A new study by Anthropic researchers finds that while robots can technically perform 74% of physical work, high costs mean only 0.3% of tasks are currently economically viable for automation.
The study runs from September 29 to October 6, 2026, and is open to Free, Pro, and Max users on Claude and Claude Code with accounts at least two weeks old.
Anthropic reports that Zhipu AI's GLM-5.3, despite having built-in safeguards, can be easily bypassed or modified to enable unrestricted cyber exploitation.
Anthropic physicists Liam Fitzpatrick and Siddharth Mishra-Sharma used the Fable 5.1 model within the Claude Science platform to compute the 9-loop MHV six-particle amplitude in planar N=4 super Yang-Mills, a result independently validated by Lance Dixon.
AI Briefing