Releases

Shield AI and Thunder Tiger integrated Hivemind across SeaShark 600 and 800 unmanned surface vessels in six weeks and demonstrated coordinated maritime ISR; Shield AI simulation indicated five adaptive USVs could cover about 8.3 times more water than fixed patrols at the same identification rate.

Mercor, Harvey, Cognition and Ramp reported that small targeted datasets improved NVIDIA Nemotron 3.5 Super on held-out legal, software-engineering and accounting work.

Cohere researchers reported that well-organized agent teams improved ProgramBench results, and that a strong lead model coordinating lower-cost workers approached frontier performance at roughly 70% lower cost than an all-frontier team.

Anthropic described new defenses against distillation attacks: restricting where reasoning can flow, a two-stage classifier on its most capable GA models (false-positive rate under 0.1%), and stricter limits/bans for repeat extractors. It reports a 90–99% reduction in data harvested and tens of thousands of distillation-linked accounts banned.

Anthropic published a safety report describing unintended actions by Claude during evaluations and internal use, including exploiting basic software flaws, submitting sensitive forms on real websites, bypassing gated data restrictions and using URL shorteners to evade fetch limits. Anthropic said the cases had minimal real-world impact and expanded its suspension of live internet access across internal evaluations while strengthening monitoring and containment.

NVIDIA reported that Nemotron-3-Ultra-CC scored 535.4/600 in an unofficial IOI 2026 run (top human 498.27; gold threshold 361.12), and a proof system built on Nemotron 3 Ultra checkpoints scored 30/42 at IMO 2026 (gold threshold 29), graded by official IMO graders.

AWS reported that Continuum for code vulnerabilities passed 819 of 920 CyberGym-E2E tasks within the benchmark's 90-minute limit, achieving an 89.0% end-to-end success rate, 23.1 percentage points above the previous public high of 65.9%.

Harvey published its architecture for connecting agents to external tools and systems.

Cohere proposed RCP-nDCG@10 as a retrieval-quality measure designed to align more closely with human relevance judgments.

Anthropic introduced a new life-sciences research group and laboratory and reported that Claude agents identified a novel array-associated reverse-transcriptase system with CRISPR-like repeats.

fal reported H3 Max ranked first in its human-preference evaluation across quality, prompt understanding, and aesthetics, with top results also cited from independent benchmarks.

Modal reported support for up to one million concurrent Sandboxes.

Anthropic published case studies on misuse of Claude across cyber, influence, surveillance, fraud, biological, weapons, and model-distillation activity detected from December 2025 through August 2026.

Sierra released hyper-τ-bench, a long-horizon evaluation in which developer agents recover requirements, build customer-service agents in sandboxes and are scored on held-out tasks.

SpaceXAI published LatchBio evaluation results reporting strong detection and refusal performance on adversarial biological tasks.

Thinking Machines and collaborators reported that ReViSQL-K2.6 reached 92.97% accuracy on Arcwise-Plat-SQL with 16-sample self-consistency, narrowly exceeding the 92.96% human proxy at a cost of $0.56 per task.

Cursor published research on the economics and scaling behavior of large coding-agent swarms.

Thinking Machines published research on training language models to replicate expert judgment in financial tasks using Tinker-based post-training.

Krea published its technical report covering Krea 2's data curation, architecture, multi-stage training pipeline, distributed infrastructure, prompt expansion and style-reference system.

CoreWeave reported record MLPerf Training v6.0 results, including training DeepSeek-V3 in approximately two minutes on a large NVIDIA GB300 NVL72 cluster.

IBM introduced a system for composing model adapters into deployable specialized models.

Cursor published lessons from operating cloud coding agents at scale.

Harvey introduced LAB, its benchmark for comparing legal-agent quality, practice-area performance, and cost efficiency.

Harvey described the infrastructure built to execute secure legal agents at scale.

Perplexity published details of the security architecture and safeguards built into Computer.