The AI Brief

Anthropic announced the Life Sciences Verification Program (LSVP) on September 17, 2026. The program gives verified life science organizations access to the Mythos 5.1, Opus, and Sonnet model families under a set of safeguards retuned to be more permissive for biology-related work. Dozens of organizations completed an early-access phase; applications are now open to the broader life science community. The program launches in beta for teams and institutions; individual Pro and Max plan access will follow later.

LSVP operates on two grant tiers: Standard Use, covering most biology research and development workflows, renewed annually; and High-risk Use, scoped to single projects and renewed every six months. High-risk grants still require coordination with the US government before Mythos 5.1 access is activated. Anthropic confirmed the safeguard shift runs from real-time blocking to offline monitoring with a 30-day data retention window. Cybersecurity safeguards remain in place across all LSVP tiers; the biology-specific relaxation is domain-scoped, not blanket. Confirmed early adopters include Xaira Therapeutics, Edison Scientific, and Manifold Bio. Access runs through the API, Enterprise, and Team plans; Claude Science, Claude.ai, and Claude Code are supported surfaces. The announcement links to Anthropic's September 2026 threat intelligence report, which documents five biological misuse case studies the company disrupted between December 2025 and August 2026, including a blocked gain-of-function grant application involving chikungunya virus research intended for a military research institute.

The LSVP is Anthropic's second domain-specific access program alongside the existing Cyber Verification Program (CVP) for cybersecurity defenders. Established partnerships with Novo Nordisk, the Allen Institute, and the Howard Hughes Medical Institute provide the institutional scaffolding. The announcement is paired with a $1 million protein design competition co-sponsored with Adaptyv Bio, offering wet-lab validation for more than 5,000 designs.

Disclosure: Claude, which generates this brief, is built by Anthropic.

Anthropic published the research post "How Claude is uplifting biomolecular modeling" on September 17, 2026. An internal general-purpose research model, operating within the Claude Science framework, produced optimized packages for 36 model implementations covering more than 30 open-source models across six families: co-folding and structure prediction (14 packages), hallucination (3), structure generation (6), inverse folding (3), genomics (7), and protein language models (3). Speedups ranged from 2.3x for OpenDDE to 6.4x for Chai-1 in fast mode. The new FlashPairformer kernels outperformed NVIDIA's cuEquivariance by 2.7x to 2.9x on triangle attention operations.

The new Big (low-memory) mode enables accurate prediction of biomolecular systems exceeding 10,000 tokens on a single NVIDIA GPU node; a token in this context represents an amino acid, nucleotide, or small-molecule atom, not a text token. Across FoldBench-Lite, 1,925 model-target pairs over 13 model configurations, prediction accuracy was unchanged: 54.8% acceptable interfaces at default versus 54.5% for Fast mode. Stated limitations include measurements taken only on H100 GPUs, and Exact and Fast modes using up to 3.2x more memory than default. All optimized code is released under Apache 2.0 but is explicitly unmaintained, with pinned upstream versions. The release is paired with a protein design competition co-sponsored with Adaptyv Bio, backed by up to $1 million in Claude credits and wet-lab validation for more than 5,000 designs.

Disclosure: Claude, which generates this brief, is built by Anthropic.

OpenAI launched the Agents API in public beta on September 10, 2026. The API brings the orchestration layer underlying Codex to developers through a flexible interface supporting long-running tasks, parallel tool calls, MCP (Model Context Protocol) integration, and web search. OpenAI-hosted sandboxes and self-hosted environment options are both available. Billing is token and tool usage only; there is no additional platform fee. Early enterprise deployments reported by OpenAI include a 60% cost reduction for one customer (SafetyKit) and an 86% failure-rate reduction for another (Hypha).

The Agents API is the developer-facing infrastructure move ahead of OpenAI's September 29 DevDay at Fort Mason in San Francisco, where the company is expected to announce Managed Agents, a higher-level no-code product layer built on the same harness. OpenAI CEO Sam Altman will deliver the opening keynote via livestream. The Agents API launch also carries a deadline: OpenAI announced in June 2026 that Agent Builder and Evals would stop working November 30, 2026. Evals go read-only on October 31. OpenAI's export tool migrates workflow structure but explicitly cannot transfer external integrations, credentials, or knowledge sources, leaving teams with a manual reconnection task before the November deadline.

OpenAI announced on June 3, 2026 that both Agent Builder and Evals would stop working at the end of November. Evals go read-only on October 31, giving teams a hard export deadline six weeks from today. Agent Builder itself shuts down November 30. OpenAI's migration export tool transfers the structural skeleton of a workflow to the Agents SDK or Managed Agents, but explicitly cannot migrate external integrations, tool credentials, or knowledge-source connections, each of which must be manually reconnected. Teams with HIPAA BAA or Zero Data Retention requirements have no confirmed clean path to Managed Agents yet, as OpenAI has not confirmed compliance eligibility for the self-hosted environment option.

The Managed Agents platform, expected to be announced formally at OpenAI DevDay 2026 on September 29 at Fort Mason in San Francisco, was found in OpenAI's codebase by TestingCatalog three weeks before the event. OpenAI CEO Sam Altman delivers the keynote via livestream. DevDay Exchanges run in parallel in eight cities: Bengaluru, Tokyo, Seoul, Paris, Berlin, London, Sao Paulo, and Mexico City. The geographic spread raises data-residency questions for any enterprise deploying on a managed runtime across those markets. First covered in Vol. I, No. 108.

NPR reported on September 18, 2026 that Google is actively participating in drafting chatbot safety bills in multiple states, including Hawaii, Iowa, Oregon, and Washington, while simultaneously lobbying those legislatures for industry-friendly carve-outs. The state bills are modeled on California's SB 243, a 2025 chatbot bill that consumer advocates have described as industry-friendly. The NPR investigation documented cases involving minors encountering sexual content on AI chatbot platforms without having initiated or invited it, which prompted the legislative push.

At the federal level, no AI-specific chatbot safety law exists. A June 2026 White House executive order announcement affirmed an intent to avoid "overly burdensome regulation" in AI. The pattern leaves state-level law as the primary near-term compliance environment for consumer-facing AI products, with the rules being actively shaped by the companies subject to them. California's SB 947, the No Robo Bosses Act, remains on Governor Newsom's desk through September 30; Connecticut's AI Responsibility and Transparency Act takes effect October 1. Operators building consumer-facing AI in the four states with active chatbot bills should monitor the exemption language, not just the headline requirements, as that is where the compliance delta between platforms will be set.