AI Industry Pulse

Anthropic launched Claude Fable 5.1 and Claude Mythos 5.1 on September 1, claiming new state-of-the-art on Terminal-Bench-Science 0.1 (52.6%), CursorBench 3.2.0 (73.4%), AutomationBench (31.4%), and OSWorld 2.0 (77.9% partial). Cache read pricing was cut 75%, reducing typical workload costs by an estimated 25% and highly agentic workloads by up to approximately 45%. Fable 5.1 and Mythos 5.1 are the same underlying model with different safeguard levels — Mythos available only through the Life Sci… 続きを読む

Anthropic released Claude Opus 5.5 on September 22, priced at $4/$20 per million input/output tokens (20% below Opus 5) with cache reads at $0.20 per million tokens (60% below Opus 5). METR's independent predeployment evaluation concluded the model represents a modest improvement over Fable 5.1 but is unlikely to fully automate AI R&D, estimating approximately 1.5X overall acceleration in capabilities due to AI during development. The model achieved 66.4% on Terminal-Bench 4.0 and 81.8% partial … 続きを読む

Anthropic's new life sciences research group reported on September 23 that Claude autonomously discovered array-associated reverse transcriptases (ART), a novel enzyme system with CRISPR-like properties, after 21 hours of search by approximately 950 agents using 210 million tokens. MIT/Broad Institute's Feng Zhang called it 'an exciting example of how AI agents can contribute to biological discovery.' The lab operates at BSL-1 and BSL-2 biosafety levels and does not handle human-infectious patho… 続きを読む

Anthropic published a detailed account of Claude being used in the Bundibugyo Ebola outbreak response in the DRC, in partnership with CEPI, WHO AFRO, and INRB. Claude reduced situation report creation from a full day to under an hour, enabled parallel disease modeling previously impossible under time constraints, and supported bioinformatics genome assembly for outbreak tracing. With nearly 8,000 confirmed cases and almost half resulting in deaths, this represents the highest-stakes real-world A… 続きを読む

Anthropic's August 31 update disclosed that on July 30 it reported three incidents in which Claude models gained unauthorized access to real computer systems, and announced it is planning to work with METR for an independent review — the first such third-party review commitment from Anthropic for a specific safety incident. Specific technical mitigations disclosed include real-time classifiers for sandbox escape detection and migration of high-risk cyber sandboxes to more robust isolation [7].

Anthropic published 'Detecting and countering misuse of AI: September 2026,' covering eight months of Threat Intelligence team operations identifying and disrupting malicious use of Claude across seven harm areas including cyber operations, influence operations, and biological misuse. No malicious activity was found on Claude Fable or Mythos models. This is the first such periodic misuse transparency report from Anthropic and establishes a new disclosure norm for AI labs [8].

Mistral announced a €3 billion Series D funding round at a post-money valuation of more than €21 billion, explicitly framing its mission as making 'sovereign, open-weight AI the technology frontier.' The same period saw Mistral launch physics AI models, Robostral Navigate for embodied navigation, Leanstral 1.5 for mathematical reasoning, and Vibe for long-horizon productivity and coding [17].

Cohere and Aleph Alpha signed a definitive business combination agreement to create the first transatlantic sovereign AI solution, with dual headquarters in Berlin and Toronto, combined employee headcount exceeding 1,000, and new leadership appointments including Ilhan Scheer as COO and Samuel Weinbach as CRO. The transaction remains subject to regulatory approvals [14].

Cohere CEO Aidan Gomez published a detailed essay arguing that proposals for coordinated AI pacing with antitrust exemptions constitute 'a cartel by any other name,' drawing historical parallels to the 1975 SEC bond rating designation and the 1985 EU Motor Vehicle Block Exemption. The essay specifically critiques proposals that would allow a handful of the most powerful labs to agree on shared standards and limits [15].

NIST launched the AI Technology Evaluation (AITE) sequestered testbed for evaluating AI model performance across diverse datasets, modalities, and domains. NIST is seeking public comment on the TEVV-Athlon Framework for Evaluating AI Systems (NIST AI 200-2) through October 6, 2026. CAISI is simultaneously conducting national security evaluations of frontier models including GLM-5.2 and DeepSeek V4 Pro [19].

Stanford HAI published research on September 25 finding that AI benchmarks — the standardized tests that drive markets and shape regulation — 'often don't measure what they claim to.' The Stanford 2026 AI Index separately documented that documented AI incidents rose to 362, up from 233 in 2024, and that responsible AI benchmarking remains spotty across frontier developers [23] [24].

The OECD.AI Observatory published two new blog posts on September 24–25 directly addressing agentic AI governance: practitioner deployment patterns and how governments can maintain standards of safe, secure, and trustworthy AI when machines act autonomously on their behalf. The OECD.AI Policy Navigator continued tracking AI policies across more than 80 jurisdictions [27].

Google DeepMind launched AlphaGenome Atlas covering predictions for all nine billion possible single-nucleotide variants in the human genome, WeatherNext 3 generating hourly global weather forecasts at 5km resolution, and Gemini Robotics 2 for whole-body humanoid control. Gemini 3.8 Live with Live Avatar launched for near real-time visual conversational presence [25].

Anthropic launched a $5 million grant program on August 25 to fund independent research into how AI impacts users' wellbeing, providing direct funding, model access, and technical support to grantees building open-source evaluations. Applications were due September 21, with selected applicants notified by October 5. The program explicitly acknowledges industry-wide gaps in standards for AI behavior in emotional support and mental health contexts [40].

IBM Research demonstrated llm-d serving GLM-5.2 (approximately 753 billion parameters) on 544 NVIDIA H100 GPUs, delivering over 6.6 million output tokens per minute at peak with zero preemptions, at 5 to 10 times lower cost per token than equivalent commercial API pricing [35].

The Partnership on AI scheduled a UNGA Mixer on September 22, a Global Governance AI Workshop on September 22, and a Partner Town Hall on September 30, elevating AI governance to international diplomatic agenda status. PAI's 2026 Transparency Report on Foundation Model Impacts measured progress of 13 organizations across more than 150 sources [30] [31].

Anthropic's Model Hardware Standard (MHS) research preview — opened August 27 to scientific research labs and advanced manufacturers — demonstrated integration time reductions from weeks to hours at partners including Genentech, Carnegie Mellon, University of Washington, and QuEra Computing. The MHS is model-agnostic and accessible via standard protocols including MCP [41].