AI Labs Unable to Guarantee Agent Safety to NYC Lawmakers; New Models Debut
Regulatory pressure on AI safety intensified as major labs introduced new models and committed significant funding to biological AI research.
The story
The question of AI agent safety took center stage this week as representatives from OpenAI, Anthropic, Google, and Meta testified before the New York City Council on October 7. During the hearing, none of the companies could guarantee that their AI agents would always comply with safety guardrails, stating that eliminating all risk or promising perfection is not feasible.
This public acknowledgment highlights ongoing challenges in controlling advanced AI systems. In a related move, Anthropic expanded access to Claude's advanced cybersecurity capabilities for vetted security professionals, allowing broader use of models with fewer safeguards for defensive work.
Meanwhile, the pace of model development continued, with new releases from Anthropic, Mistral, and Google, alongside a substantial $1.8 billion investment into foundational biological data for AI models from a consortium including Google DeepMind and Meta. These parallel developments underscore both the rapid advancements in AI capabilities and the growing scrutiny over their potential risks and ethical deployment.
Who moved
Anthropic
What Changed: The company released Claude Haiku 5.5 and expanded access to Claude's advanced cyber capabilities for vetted security professionals.
Consequence: Claude Haiku 5.5 is priced competitively at $0.10/$0.50 per Mtok for prompts up to 100,000 tokens, while expanded cyber access allows security professionals broader use of models with fewer safeguards for defensive operations.
Mistral
What Changed: Mistral opened a public preview of its Large 4 model, nicknamed Le Chonk, via API.
Consequence: The 1.05-trillion-parameter model is priced at $0.68/$2.09 per Mtok, with open weights promised by the end of October, expanding accessible large language model options.
What Changed: Google released Nano Banana 2.1, deprecating Nano Banana 2, and participated in a major biological AI funding initiative.
Consequence: Nano Banana 2.1 offers new pricing for image processing in the Gemini API, while the $1.8 billion Biohub initiative, with Google DeepMind's involvement, aims to generate AI-ready biological data for disease prediction and treatment.
Meta
What Changed: Meta joined Google DeepMind and Isomorphic Labs in contributing $300 million to the Virtual Biology Initiative.
Consequence: This investment is part of a larger $1.8 billion international effort to create technologies and multi-modal datasets for predictive AI models in biology.
New models
Claude Haiku 5.5
Lab: Anthropic
What: A new model priced at $0.10 input and $0.50 output per million tokens for prompts up to 100,000 tokens.
Use: It offers a cost-effective option for users requiring high-capacity language model capabilities.
Mistral Large 4 (Le Chonk)
Lab: Mistral
What: A 1.05-trillion-parameter mixture-of-experts model with 49 billion active parameters, capable of reading text and images.
Use: Available as a public preview API, it provides a powerful, multimodal option for developers, with open weights expected by month-end.
Nano Banana 2.1
Lab: Google
What: The direct successor to Nano Banana 2 (Gemini 3.1 Flash Image), with updated pricing and features for image processing.
Use: It is available in the Gemini app, Google Search AI Mode, Google AI Studio, Flow, and the Gemini API for image-related tasks.
d1-3B and d1-omni-600M
Lab: LiquidAI
What: Two new model variants released with open weights under an LFM Open License v1.0.
Use: These models provide additional open-source options for developers and researchers.
pplx-embed-v2-late-0.6b and 9b
Lab: Perplexity
What: Two new multimodal embedding models released with MIT license weights.
Use: They offer tools for developers to integrate Perplexity's embedding capabilities into their applications.
Market signals
Crunchbase reported on October 7 that North American startup funding in Q3 2026 totaled $92 billion, with AI companies securing $61 billion.
Implication: AI-focused companies continue to attract significant capital, capturing roughly two-thirds of total investment, despite an overall quarterly decline in funding attributed to the absence of new mega-rounds for some top labs.
Biohub, the U.S. Department of Energy, and the National Institutes of Health, along with Google DeepMind, Isomorphic Labs, and Meta, announced a $1.8 billion international initiative on October 7.
Implication: This coordinated commitment, including $300 million from Google DeepMind, Isomorphic Labs, and Meta for the Virtual Biology Initiative, signals a major push by labs into foundational biological data generation for AI-driven disease prediction and treatment.
OpenAI, Anthropic, Google, and Meta representatives testified before the New York City Council on October 7, unable to guarantee their AI agents would always comply with safety guardrails.
Implication: The public hearing highlights increasing regulatory scrutiny on AI safety and the inherent challenges in ensuring absolute control over advanced AI systems, potentially leading to new legislative proposals.
What we'll be watching
- Mistral plans to release open weights for its Large 4 model by the end of October.
- Andreessen Horowitz's San Francisco Tech Week continues through October 11.
- The Open Compute Project's OCP Global Summit is scheduled for October 12-15 in San Jose.
- Anthropic is targeting a potential initial public offering (IPO) launch in October.
- Danaher's first AI-powered autonomous lab is expected to operate at scale in early 2027.
Reporting + analyst voices: grounded via Google Search at publish time.