The Daily Brief · Issue 11 · 28 August 2026Judge blocks the Pentagon's Anthropic blacklist as two Chinese labs publish giant open-weight models
A federal judge called the Defense Department's move against Anthropic illegal. Tencent and Z.ai both put frontier-class weights online, and Anthropic, Google and others shipped tools for scientists and video makers.
RegulationFederal judge blocks the Pentagon's blacklisting of Anthropic
On August 27, US District Judge Rita Lin blocked the Defense Department's designation of Anthropic as a national-security supply-chain risk. In a 59-page order she called the decision illegal and baseless, and wrote that national security is not a blank check to punish critics. Defense Secretary Pete Hegseth had applied the rarely used procurement rule in February 2026, after Anthropic refused to let Claude be used for domestic surveillance or autonomous weapons. Anthropic argued the move broke its free-speech and due-process rights and could cost it billions. A second Anthropic lawsuit in Washington, D.C., over a separate designation that could shut it out of civilian government contracts, is still pending.
Why it matters: Government buyers can again consider Claude for defence work while the case continues, but the civilian-contract fight is unresolved.
Source: NBC News
Open weightsTencent open-sources Hy4 preview, a 770B model under Apache 2.0
Tencent released Hy4 preview on August 28 with open weights on Hugging Face, ModelScope, GitCode and CNB. It is a mixture-of-experts (MoE) model: 770 billion parameters in total, but only 49 billion are used for each token, which keeps it cheaper to run. It reads up to 1 million tokens of context. Tencent reports 92.3 on GPQA Diamond (graduate-level science questions) and 65.7 on SWE-bench Pro (real coding tasks). Those are the company's own numbers, not independent tests yet. The licence is plain Apache 2.0, with no user caps or field-of-use limits, so companies can fine-tune and sell products on it. An FP8 version for smaller hardware ships alongside.
Why it matters: It is one of the largest models ever released under a fully permissive licence, raising the bar for open-source LLM rankings.
Source: Hugging Face (Tencent model card)
Open weightsZ.ai releases GLM-5.3 weights after a two-week safety hold
Z.ai (Zhipu) put the full GLM-5.3 weights on Hugging Face on August 28, two weeks after the model launched on its API. It is a roughly 753-billion-parameter MoE model with a 1-million-token context. Z.ai says it held the weights back for extra security testing because the model is strong at finding and exploiting software bugs: it reports 84.5 on CyberGym and 28.3 on Terminal Bench 3.0. The licence is not MIT. It works like MIT for most users, but any company with more than $10 billion in yearly revenue must pass a Z.ai security review before commercial use.
Why it matters: Big cloud providers cannot simply host GLM-5.3 without Z.ai's sign-off, a new twist on 'open' model licensing.
Source: Hugging Face (Z.ai model card)
PricingAnthropic offers 10,000 free or cheap Claude seats to scientists
Anthropic launched a Claude team plan for scientists on August 27. It covers 10,000 subscriptions for one year: standard seats are free, and premium seats with five times the usage limits cost $15 a month. Applicants must be a principal investigator (a lab or project lead) at an academic or nonprofit institution. Anthropic says it plans to grow the program well past 10,000 seats. It also widened its AI for Science program beyond biology, offering up to $50,000 in API credits per project. Biology and chemistry users stay limited to Opus-class models, because Fable models still block professional biology and drug-development questions over misuse risk.
Why it matters: University researchers can now get paid-tier Claude for free, which changes the maths in our research-tool rankings.
Source: Anthropic
AgentsAnthropic previews a standard for AI agents that run lab hardware
Anthropic published a research preview of the Model Hardware Standard (MHS) on August 27. It is a shared specification that lets AI agents safely control physical devices such as pipetting robots and lasers. Anthropic says wiring an agent to new hardware usually takes weeks or months, and MHS aims to cut that to hours. Partner results are vendor-reported: Carnegie Mellon cut one integration from several weeks to about 8 hours, and QuEra raised its laser recovery success rate from 58% to 99.3% while cutting recovery time from 150 seconds to 6. Genentech and Tetsuwan Scientific also tested it on lab workflows.
Why it matters: If other labs adopt it, agents could move from screens into real lab equipment much faster.
Source: Anthropic
LaunchGoogle makes Gemini Omni 1.1 Flash video model generally available
Google moved gemini-omni-1.1-flash to general availability in the Gemini API on August 27. GA means it is now a stable, supported product rather than a preview. The video model adds scene extension, first-and-last-frame interpolation (you give the start and end frames and it fills in the motion), video references, fast 360p drafts and upscaling to 4K. It is rolling out in Google AI Studio, Flow and the Gemini app. Developers on the old preview endpoint need to switch, because Google will shut that endpoint on September 30, 2026. Google has not changed its headline pricing in the release note.
Why it matters: Developers building video apps get a supported Google model with editing controls, and have one month to migrate off the preview.
Source: Google AI for Developers (Gemini API changelog)