The AI Brief

Meta's Superintelligence Labs released Muse Glimmer on Monday morning, posting weights on Hugging Face under an Apache 2.0 license. Muse Glimmer is a 30-billion-parameter model optimized for always-on local agent workflows, small enough to run on a Mac or PC with a single consumer GPU, enabling local agents, function calling, local coding, and LLM-as-a-judge evaluation. Meta said users will be able to download and customize Glimmer, a distilled version of the company's Muse Spark 1.2 model designed with a focus on efficiency to minimize system requirements.

Meta compares Muse Glimmer with Google's Gemma and Alibaba's Qwen rather than with the frontier giants, positioning it as a leader in its size class and a bet that the future is not only enormous cloud systems but also smaller models that live on your own machine. Meta also plans to release the weights of Muse Spark 1.2, its most advanced model, built by a costly superintelligence team it formed last year.

Zuckerberg's policy argument was as prominent as the model itself. Zuckerberg said "Foreign labs currently hold several advantages here since American labs have to comply with many additional restrictions on training data," and that "US policy must reduce this additional friction if we want American open source models to lead over time." In a video post accompanying his 14-page essay titled "The Future is for Everyone," Zuckerberg said "We've got even bigger models that are coming soon." Chinese startups are leading the race for open-weight models, with Moonshot's Kimi K3, Alibaba's Qwen3.8-Max, and DeepSeek's V4-Flash delivering performance that rivals top systems by US AI labs, while the leading models of US developers OpenAI, Anthropic, and Alphabet's Google remain closed source. Zuckerberg also unveiled a new $1 billion fund aimed at easing community opposition to Meta data-center construction.

OpenAI's confidential S-1 was submitted to the SEC on June 8, 2026, and as of August 9, 2026, the full public prospectus has not yet appeared on EDGAR. The public S-1 prospectus is expected on SEC EDGAR in mid-to-late August 2026, roughly 15 days before any roadshow. The filing will disclose audited financials, the Microsoft revenue-share agreement, and detailed risk factors for the first time.

OpenAI is generating approximately $2 billion in revenue per month but remains unprofitable, reporting losses of roughly $1.22 for every $1 earned; the company's most recent private-round valuation stands at $852 billion, and a public listing could target a valuation above $1 trillion. Internal documents suggest management is projecting a $14 billion loss in 2026 and that the company does not expect to be profitable until 2029. Goldman Sachs and Morgan Stanley are leading the filing process.

The OpenAI Foundation, the nonprofit, retains board-appointment control over OpenAI Group PBC, meaning public shareholders will not have standard governance power over the company they invest in. A September or Q4 2026 listing remains the target, though a 2027 debut is also under consideration.

Claude Sonnet 5 intro pricing ends in 31 days from early August: $2/million tokens becomes $3/million on September 1, plus a tokenizer that adds up to 35% more tokens per equivalent text. The change applies to API usage; plan-included access terms may differ.

The tokenizer shift is the less-discussed mechanism. When Anthropic updates the tokenizer alongside a price-tier change, the per-token price and the token count both move, operators who stress-tested cost models against the $2/million intro rate with the existing tokenizer will need to rerun those estimates against both variables. Teams using Sonnet 5 for long-document processing or high-frequency short queries are most exposed, since tokenizer inflation is not uniform across task types.

Anthropic's current lineup, Haiku 4.5 for cost, Sonnet 5 for balance, Opus 5 for capability, means operators have tiering options. Opus 5 launched July 24 at $5/$25 per million tokens and currently leads the Artificial Analysis Agentic Index, while Sonnet 5 has served as the mid-tier workhorse for teams that do not need full frontier capability.

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

First covered in Vol. I, No. 76. As of August 8, 2026, Gemini 3.5 Pro remains in limited preview on Vertex AI and has not launched publicly; at Google I/O Sundar Pichai said it was in internal use, the previously rumored July launch date has since passed without a release, and a new rumor now points to August 12, still unconfirmed by Google.

Google will hold its Made by Google event in New York City on August 12, where the company plans to introduce its Pixel 11 smartphone series. Kalshi data shows traders have wagered more than $320,000 on when Google will release Gemini 3.5 Pro; traders assign a 46% chance to a release before August 16, rising to 63% before August 21 and 76% before August 31. Google has not said whether Gemini 3.5 Pro will appear at the Pixel event and has not confirmed the model's features, pricing, access plans, or rollout schedule.

A Bloomberg report had revealed that the launch was delayed by months as engineers worked to improve the model's capabilities, particularly in coding. Leaked benchmarks cited by community sources indicate improvements in SVG generation, frontend coding, and multi-step agentic workflows, but no official model card has been published. The delay has accumulated across five missed targets since Google I/O in May.

Anthropic launched Claude for Government Desktop in public beta on July 7, 2026, bringing Claude Code and Claude Cowork to public sector agencies in a FedRAMP High authorized environment. The product packages Claude Code and Claude Cowork for US federal, state, and local government agencies as a desktop application that deploys through standard agency MDM platforms without requiring a separate cloud-provider contract.

It is a distinct desktop application running inference inside a FedRAMP High authorized environment; conversation history and file access stay on agency-managed devices, and the governance layer is substantially heavier than the commercial product. The beta adds desktop file-based work, stronger admin controls, tamper-evident audit logs, and spend governance for agencies.

A limited-time program makes Claude for Government available to federal agencies at $1 per agency with unlimited seats through August 2026. Anthropic remains the contracted and billing party, agencies do not need a separate cloud-provider relationship to get started. Classified workloads still require Tier 1 or 2, using Claude Gov models on AWS.

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