The AI Brief

The Trump administration filed a 20-page statement of interest on September 2 in Manhattan federal court, backing OpenAI in The New York Times's lawsuit over training data. The brief, filed with US District Judge Sidney Stein in the Southern District of New York, calls AI training on publicly available internet material fair use, describing it as "extraordinarily" transformative and arguing that constraining LLM development under a "misunderstanding" of fair use doctrine would threaten scientific progress and US economic competitiveness. It is the first time the US government has formally stated a position in the broad wave of copyright cases against AI developers.

The Times's suit, first filed in 2023, accuses OpenAI and Microsoft of using millions of newspaper articles without permission. The government's brief does not decide the case or bind the court, but it gives OpenAI a high-profile ally in the fair-use dispute. The brief also covers national-security framing, stating that the United States "has a strong interest in this court rejecting any argument that training LLMs on copyrighted texts violates copyright law." The same fair-use question sits at the center of separately active cases involving Sony, Warner Chappell, book authors, and other content owners across multiple courts, including Sony and Warner's suit against Anthropic filed just last week. The brief's legal weight is advisory, but its political weight accrues to all defendants simultaneously.

Google released Gemini 3.8 Flash on September 2, 2026, generally available via the Gemini API and Google AI Studio under model ID gemini-3.8-flash. The model scores 90.8% on Terminal-Bench 2.1 (up from 81.6% for 3.7 Flash three weeks prior), 73.7% on DeepSWE v1.1 (essentially tied with Opus 5's 74.0%), and 61.4% on the Vals Finance Agent v2 benchmark, leading both Opus 5 and GPT-5.6 Sol on the finance task. Output speed runs at approximately 305 tokens per second, the fastest independently clocked for a generally available model. The introductory price of $0.75/$3.75 per million tokens is identical to 3.7 Flash; Google has confirmed the standard rate of $1.50/$7.50 per million tokens takes effect January 1, 2027. Gains are uneven: HLE-Verified held essentially flat at 54.9% versus 45.7% for the prior model, indicating limited improvement on general expert-level reasoning.

Alongside the standard model, Google launched Gemini 3.8 Flash Cyber, a variant tuned for vulnerability detection and automated patching, available only through the newly announced Fairwind Program. The Fairwind Program grants gated access to government agencies, national cyber authorities, critical-infrastructure operators, and core technology platform maintainers; more than 650 organizations are participating globally at launch. Fairwind participants receive Cyber paired with CodeMender, Google's AI agent for vulnerability remediation, and must restrict model access to employees in cybersecurity, incident response, or penetration testing roles. On the CWE-Bench patching benchmark the Cyber variant posts a 47.2% pass@1 score and exceeds a 70% real-world vulnerability discovery rate. The Fairwind structure mirrors the gated access model OpenAI used for Daybreak Red, a restricted tier for high-capability security-relevant AI that does not appear in general APIs.

Anthropic's Claude Mythos 5.1, released September 1 alongside Fable 5.1, produced protein binder designs across 12 drug-relevant targets and achieved a nearly 50% hit rate, experimentally validated by two independent organizations, Adaptyv Bio and a second external firm. The 10–15% hit rate is the current field norm Anthropic cites from public data in the Proteinbase database. On three specific competition targets, EGFR, Nipah G, and 15-PGDH, drawn from Adaptyv Bio's open public design competitions, Mythos 5.1's binding affinities ran ten times higher than the best designs submitted by human competitors. Every design shown in Anthropic's published video was confirmed to bind in the lab.

The compute budget is substantial: Anthropic gave models up to 12,500 Nvidia H100 hours over a 48-hour session in multi-target mode, and up to 2,500 H100 hours per target in single-target mode. Single-target mode produced the strongest results. The campaign used open-source protein design and folding tools rather than proprietary biology infrastructure, a detail that matters for reproducibility. Anthropic is enrolling external life-sciences organizations in a gated access program for Mythos in partnership with the US government, and plans to expand access to the broader research community. The prior protein binder result disclosed in August used Mythos Preview and Opus 4.8; these results are Mythos 5.1, a materially upgraded model on the same task.

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

OpenAI launched the Epic EHR integration and Healthcare Public Data plugin for ChatGPT for Healthcare on September 1, 2026. The Epic plugin gives clinicians read-only access to authorized patient information from their organization's Epic record, reviewing history, identifying pending screenings, and preparing for appointments, requiring an administrator-configured Epic EHR app, individual Epic sign-in, and existing patient-chart permissions. In supported deployments, ChatGPT embeds directly inside the EHR layout, meaning clinicians access AI assistance without navigating away from the patient chart. UCSF Health served as a pilot partner. The integration covers Epic's installed base of over 325 million patient records.

The Healthcare Public Data plugin adds structured access to nine official sources including ClinicalTrials.gov, CMS Coverage, RxNorm, DailyMed, and PubMed, all read-only, none accessing patient charts. Both plugins are available to HIPAA-enabled ChatGPT Enterprise and ChatGPT for Healthcare workspaces. The release runs on the same BAA-governed, role-based-access and audit-log infrastructure OpenAI introduced for regulated workspaces. The release also makes the same plugins available inside Codex, OpenAI's coding agent, for eligible workspaces, an extension that is narrow today but positions Codex for clinical data engineering use cases.

Google launched the Fairwind Program on September 2 alongside Gemini 3.8 Flash Cyber, giving more than 650 organizations globally application-gated access to the cybersecurity model and CodeMender, its autonomous vulnerability-remediation agent. Participating organizations must restrict access to employees in cybersecurity, incident response, or penetration testing functions; deploy multi-factor authentication; and operate within controlled cloud environments. Priority goes to government agencies, national cyber authorities, critical-infrastructure operators, healthcare networks, energy grids, financial systems, and maintainers of widely used open-source software. The model itself is not available through the Gemini API or Google AI Studio.

The structural parallel is precise. OpenAI's Daybreak Red tier for GPT-5.6-Cyber, announced in August, routes through a vetted enterprise security program with zero-operator-access enforced at the chip layer. Anthropic's Mythos gated life-sciences access program, announced alongside Mythos 5.1, uses government partnership for enrollment. All three programs restrict access by organization type and role rather than by price, a departure from the standard tiered-pricing model that has governed API access since GPT-3. The implication for security teams: the ROI case for investing in the vetting process and operational controls required for program eligibility now includes access to model capabilities that are not purchasable at any price through standard channels, making program membership a procurement category of its own.