The AI Brief

Trial Lawyers for Justice filed the complaint on September 19 in the U.S. District Court for the Northern District of California on behalf of four paying subscribers to the defendants' AI services. The suit names Anthropic PBC, OpenAI OPCO LLC, Google LLC, and SpaceXAI as defendants and alleges a violation of Section 1 of the Sherman Antitrust Act, 15 U.S.C. §1, which prohibits agreements among competitors to restrain trade. The plaintiffs seek class certification, injunctive relief, and damages.

The complaint centers on September 12, when Anthropic CEO Dario Amodei published "We Must Pace the Frontier," calling for industry-wide coordination to decelerate AI capability development. That same day, OpenAI CEO Sam Altman publicly endorsed the proposal, SpaceXAI CEO Elon Musk said "Dario is right," and Google DeepMind co-founder and chair Demis Hassabis called the essay "the right path forward." The plaintiffs argue those responses constituted an agreement to restrict output, and that the coordination may have begun as early as July 2026 when employees from multiple labs signed a joint statement. Lead attorney Nick Rowley argued antitrust law does not permit competitors to collectively decide that competition is too dangerous.

Amodei had flagged the antitrust exposure in the original essay, suggesting Congress could provide a "narrow waiver for certain safety conversations." Senator Josh Hawley rejected that idea during a September Senate hearing, calling any antitrust exemption for the largest technology companies unacceptable. President Trump has also rejected the pacing calls. Representatives for all four defendants had not responded to comment requests as of Sunday. The filing means every future public statement by a lab CEO on pacing is now potential evidence in active litigation.

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

Alibaba's Qwen team released Qwen3.8-Omni-Flash on September 18 through QwenCloud, Alibaba Cloud Model Studio, and Qwen Chat, with availability confirmed in Beijing, Singapore, Hong Kong, Tokyo, Frankfurt, and Virginia. The model is API-only at launch; no open weights were announced. It accepts text, images, audio up to three hours, and video up to two hours and 2 GB via URL, all within a single 1M-token context window, with text-only output and native function calling, web search, and structured outputs. Audio input spans 113 languages and dialects.

Pricing is listed at $0.15 per million input tokens and $0.47 per million output tokens on QwenCloud, with cached inputs at $0.016 per million. Alibaba reports audio-visual combined input costs 93% less than Qwen3.5-Omni-Plus and video input roughly 89% less. The model achieves a 26% average improvement across 30 vendor-run evaluations versus Qwen3.5-Omni-Plus. Efficiency gains trace partly to 51.8% fewer tokens on OmniVideoBench at comparable accuracy. All benchmark figures are Alibaba's own; no independent evaluation existed at publication. Third-party comparisons with Gemini 3.8 Flash show benchmark parity on multimodal tasks, with Qwen priced approximately 80% lower on input and 88% lower on output.

Caltech spinout PrismML released Ternary Bonsai 2 27B on September 17 under Apache 2.0, available on Hugging Face for free download. The model is derived from Alibaba's Qwen3.8 27B open-weight base and uses ternary weights with FP16 group-wise scaling, yielding a total footprint of 5.9 GB. PrismML was founded with support from Khosla Ventures, Cerberus, and Google, with continuing backing from Samsung. UC Berkeley professor and advisor Ion Stoica said the compression gap closing to 1.8% "can fundamentally expand where highly capable models can be deployed."

On a 20-benchmark aggregate spanning reasoning, math, coding, instruction following, vision, and agentic tool use, Bonsai 2 scores 83.9, retaining 98.2% of Qwen3.8 27B's 85.4. The losses concentrate in long-horizon tasks: SWE-bench Verified drops from 80.6 to 60.8, and Terminal-Bench 2.1 drops from 69.7 to 52.8, roughly 75% retention on both. Math and coding retention is near-complete. On an NVIDIA RTX 5090, the model runs at 142.5 tokens per second; on an Apple M5 Max laptop, at 46.8 tokens per second. The first Bonsai 27B, released in July 2026, retained only 95% of its base model's performance; the 3.2-point gain to 98.2% in two months reflects rapid iteration in low-bit training methods.

All benchmark figures are PrismML's own, run with EvalScope and vLLM on an NVIDIA H100 in thinking mode. No independent reproduction had been published at time of writing.

AIR Security disclosed Plugin4Shell on September 18, documenting a zero-click RCE that defeats SHA-pinned plugin verification in Claude Code, OpenAI Codex, GitHub Copilot, and Gemini CLI. The flaw exploits an incorrect assumption about what a git checkout guarantees in the plugin verification chain, and AIR found four independent implementations of the same wrong assumption across the four products. Anthropic shipped a fix in Claude Code 2.1.179; OpenAI shipped a fix in Codex 0.146.0. As of Sunday, GitHub Copilot remains unpatched and Microsoft has not published a patch timeline. Google announced it is deprecating Gemini CLI entirely and directing users to the Antigravity IDE rather than issuing a security fix.

The divergence in vendor response is operationally significant. GitHub Copilot is the largest installed-base coding agent by user count. The plugin auto-update behavior that makes Plugin4Shell zero-click cannot be disabled without also disabling plugin functionality. Security teams at organizations running Copilot face a binary choice pending a Microsoft patch. First covered in Vol. I, No. 118.

Beijing-based Naive AI raised $400 million across three rounds ($100M, $180M, $120M) at a $1.42 billion valuation, according to The Information citing a person with direct knowledge. Naive AI has not confirmed the figures. Investors include Tencent, IDG Capital, MPCi, and HSG (formerly Sequoia Capital China). The company was founded in February 2026 by Dai Jifeng, an associate professor in Tsinghua University's Department of Electronic Engineering whose decade of research spans deep learning, computer vision, foundation models, and agentic AI.

Naive AI is not pretraining a model from scratch. Instead, it takes an existing Chinese open-weight model as a base (unconfirmed which one) and applies mid-training, post-training, and reinforcement learning to produce a differentiated system. The company plans to release its first open-weight model as early as September 2026. The approach mirrors PrismML's Bonsai series in philosophy but targets a different part of the improvement curve: where PrismML compresses for deployment efficiency, Naive AI targets capability improvement via post-training on top of open-weight foundations. A founding dispute with prior backer MiroMind over intellectual property remains unresolved, according to The Information.