## Sources 1. [42 State AGs Probe OpenAI Days After IPO Filing](https://awesomeagents.ai/news/openai-42-states-investigation-ipo/) 2. [US Export Order Forces Global Fable 5, Mythos 5 Shutdown](https://awesomeagents.ai/news/us-export-ban-fable-5-mythos-shutdown/) 3. [Kimi K2.7-Code](https://awesomeagents.ai/models/kimi-k2-7-code/) 4. [Kimi K2.7-Code - Moonshot's Open-Weight Coding Leap](https://awesomeagents.ai/news/moonshot-kimi-k2-7-code/) --- ### **42 State AGs Probe OpenAI Days After IPO Filing by Elena Marchetti** * **Subpoena and Timing:** A coalition of **42 state attorneys general**, led by New York’s Letitia James, served OpenAI with a sweeping subpoena on **June 12, 2026** [1, 2]. This legal action occurred just **five days after OpenAI filed confidentially for a $1 trillion IPO**, making it a "material legal risk" that must now be disclosed in the company's S-1 prospectus [1, 3]. * **Core Investigative Areas:** The probe targets six distinct aspects of OpenAI’s operations: **advertising and user engagement** (tactics for retention), **consumer and health data** (storage and use of sensitive information), **protections for minors and seniors**, **internal company policies** regarding safety and governance, and **model sycophancy** [2, 4]. * **The Sycophancy Argument:** Regulators are treating **model sycophancy**—the tendency of AI to agree with users regardless of accuracy—as a **consumer protection issue** [5]. The legal theory suggests that if a model is designed to maximize engagement by telling users what they want to hear rather than what is true or safe, it may constitute a **duty-of-care problem** [5]. * **Historical Context and Escalation:** This investigation follows a **"first-in-the-nation" lawsuit filed by Florida** on June 1, 2026, which uniquely names CEO **Sam Altman personally** as a defendant [6]. The Florida complaint alleges that ChatGPT "aided and abetted" a mass shooter and lacks effective parental controls [7, 8]. * **Federal vs. State Conflict:** The probe arrives amidst a political battle over **federal preemption**, where Congress is debating whether federal AI legislation should block states from imposing AI-specific design mandates [9]. The New York subpoena challenges this boundary by demanding scrutiny of internal model design and policies [9, 10]. ### **Kimi K2.7-Code - Moonshot's Open-Weight Coding Leap by Elena Marchetti** * **Model Release:** Moonshot AI launched **Kimi K2.7-Code** on June 12, 2026, as an update to its open-weight K2.6 model [11, 12]. It utilizes a **1T-parameter Mixture-of-Experts (MoE)** architecture with 32B active parameters per token [12, 13]. * **Efficiency Gains:** The most significant technical improvement is a **30% reduction in reasoning tokens** compared to the previous version [12, 14]. For developers running agentic loops at scale, this "thinking compression" makes the model roughly **30% cheaper per task** [14, 15]. * **Benchmark Performance Claims:** Moonshot reported a **21.8% gain on its internal Kimi Code Bench v2** [12, 16]. However, the source notes that K2.7-Code still **trails Claude Opus 4.8 and GPT-5.5** on five out of six benchmarks provided in the company's own comparison table [16, 17]. * **The Transparency Gap:** A major takeaway is the **absence of independent evaluations**; all six benchmarks cited in the release are proprietary to Moonshot [18]. There are no scores for public standards like SWE-Bench Pro or LiveCodeBench, which complicates head-to-head comparisons for teams deciding between models based on task quality [18, 19]. * **Pricing and Deployment:** Reasoning tokens are billed as output at **$4.00 per million**, while input tokens cost **$0.95 per million** (dropping to $0.19 for cached input) [14, 15]. This is significantly cheaper than Claude Opus 4.8, which charges $25.00 per million output tokens [15]. ### **Kimi K2.7-Code by James Kowalski** * **Specialized Focus:** This source emphasizes that K2.7-Code is a **narrowly focused coding successor** to K2.6, with training compute directed toward **end-to-end coding tasks**, **MCP (Model Context Protocol) tool-call chains**, and instruction-following [20, 21]. * **Technical Constraints:** Unlike some competitors, **"thinking mode" is mandatory** and cannot be disabled in the API [22, 23]. Additionally, server-side sampling is **locked at temperature 1.0 and top_p 0.95**, offering no override for developers who require deterministic outputs [22, 23]. * **Tool-Use Capabilities:** One of the model's primary strengths is its **81.1 score on the MCP Mark Verified benchmark**, which tests human-verified tool use in environments like GitHub, Postgres, and Playwright [24, 25]. This score reportedly **beats Claude Opus 4.8** in the same category [24, 26]. * **Deployment Requirements:** Self-hosting the model requires substantial resources, including roughly **595 GB of disk space** [22, 27]. While native **INT4 quantization** helps reduce VRAM needs, it remains limited to well-resourced teams using multi-GPU H100 setups [23, 27]. * **Economic Advantage:** Despite trailing GPT-5.5 on agentic benchmarks by 4-17 points, K2.7-Code is positioned as the **economical choice** for high-volume workflows due to its lower token rates and improved reasoning efficiency [24, 26, 28]. ### **US Export Order Forces Global Fable 5, Mythos 5 Shutdown by Sophie Zhang** * **Unprecedented Government Action:** On **June 12, 2026**, US Commerce Secretary Howard Lutnick issued the **first-ever export control directive** against a commercial large language model [29, 30]. The order forced Anthropic to **disable Claude Fable 5 and Claude Mythos 5 worldwide** [30, 31]. * **National Security Concerns:** The directive was based on a claim that a **"narrow, non-universal jailbreak"** was discovered that could allow the models to be used for **software vulnerability scanning** [31, 32]. Anthropic noted that Fable 5 was only **three days old** when the shutdown occurred [33]. * **Global Impact and Reasoning:** Because Anthropic cannot verify the nationality of every API user in real-time, it had to shut down the models globally to ensure no "foreign national" gained access, as mandated by the order [30, 33]. * **Anthropic’s Counterarguments:** Anthropic argued that the capability to scan for vulnerabilities is **widely available in other models**, specifically citing **OpenAI's GPT-5.5**, which remains fully accessible [34, 35]. They characterized the directive as a **"misunderstanding"** and warned that applying such a standard industry-wide would halt all frontier model deployments [36, 37]. * **Operational and Financial Risk:** The shutdown provided **no advance notice or SLA protection**, causing immediate failures for integrated enterprise systems [38, 39]. This event is considered a **material risk** for Anthropic's expected IPO, as underwriters will require clarity on the probability of the directive expanding to other models [37, 40].