Sia Reads
What Matters

Zhipu AI (Z.ai) released GLM-5.2 under the MIT license, trained on ~100,000 Huawei Ascend 910B processors with no Nvidia hardware involved. On extended coding-agent trajectories, it trails Opus 4.8 by just 1% while edging past GPT-5.5. At $1.40/$4.40 per million tokens via OpenRouter, it's a fraction of frontier pricing. The timing — released one day after the US banned Fable 5 exports — makes the geopolitical subtext hard to miss.

OpenAI unveiled the GPT-5.6 series on June 26: Sol (flagship, $5/$30 per M tokens), Terra (2x cheaper than GPT-5.5), and Luna (cheapest tier at $1/$6). Sol hits 88.8% on Terminal-Bench 2.1, beating Fable 5's 83.4%. A new 'ultra' mode dispatches subagents for complex tasks. The catch: at the US government's request, it's limited to ~20 vetted partners for now, with broader access expected in the coming weeks.

Claude Science launched June 30 as a research environment integrating 60+ curated connectors for genomics, proteomics, structural biology, and cheminformatics. It renders 3D protein structures natively, manages compute across local machines and HPC clusters, and includes a reviewer agent that checks citations and calculations. Anthropic also announced an internal drug discovery program targeting neglected diseases, with Novo Nordisk and the Allen Institute as early partners.

The MCP 2026-07-28 release candidate, locked May 21, removes the initialize/initialized handshake, drops the Mcp-Session-Id header, and deprecates Roots, Sampling, and Logging. New additions include the Extensions framework, MCP Apps (server-rendered HTML UIs), and a redesigned Tasks extension. The practical win: remote MCP servers can now run behind a round-robin load balancer with no sticky sessions or shared session stores.

The Financial Times reported that Altman has been in talks with Trump, Treasury Secretary Bessent, and Commerce Secretary Lutnick about giving the US government a 5% stake in OpenAI — roughly $42.6B at current valuation. The proposal would extend to Google, Anthropic, xAI, and Meta. Altman also published an FT op-ed calling for a US-led international AI standards forum modeled on the IAEA. The discussions remain conceptual and early-stage.

Bloomberg and The Information report that Anthropic is in early talks with Samsung to manufacture a custom AI accelerator, evaluating Samsung's 2nm process and advanced packaging. The company hired Clive Chan from OpenAI's chip team in June — the second hardware engineer on what appears to be a growing silicon effort. Samsung invested in Anthropic's $65B Series H round in May. Anthropic says its compute strategy will continue to span Google, Amazon, and Nvidia hardware.

The US Department of Commerce removed export controls on Claude Fable 5 and Mythos 5 effective June 30, restoring global access from July 1. The ban was triggered after Amazon researchers discovered a safeguard bypass; Anthropic deployed an improved classifier blocking the technique in 99%+ of attempts. As part of the resolution, Anthropic committed to expanded government collaboration and is co-developing an industry-wide jailbreak severity scoring framework with Amazon, Microsoft, and Google.