The 2026-09-23 YZ Index Smoke quick test covered 10 models, with Claude Sonnet 4.6 leading at 96.98. The brief highlights daily leaderboard signals, notable score changes, and items that require follow-up runs to confirm.
Xiaomi has officially released and open-sourced the MiMo-V2.6 series, whose Pro version scored 46 on the Artificial Analysis Intelligence Index v4.3.2, surpassing Kimi K3 and GLM-5.3 to become the highest-ranked open-weight model on the current AA Index.
Alibaba's T-Head unveiled the Zhenwu V900 AI chip, claiming triple the compute of its predecessor and support for clusters scaling to 500,000 cards, but the full-stack Panjiu supernode server will not enter mass production until Q1 2027. T-Head also previewed the Zhenwu J900, based on an in-house parallel computing architecture, for March 2028.
OpenAI says GPT-6 Astra is the first model to reach the Critical cybersecurity level under its Preparedness Framework, scoring 100% on ExploitBench while showing a 0% unauthorized scope-violation rate. The release pairs unprecedented offensive capability with strict access controls, raising questions about the boundaries of responsible release.
On September 18, 2026, independent evaluator Robocurve released the RoboHarm benchmark, which found that GPT-6 Astra refused only 2 of 100 real robotic-arm dangerous-command trials, compared with 20 refusals by Claude Fable 5.1 and zero by MolmoAct2. The results highlight the limitations of existing alignment methods in embodied scenarios.
U.S. Treasury Secretary Scott Bessent told lawmakers that frontier AI labs will not receive any liability exemption or blank check, arguing that holding creators responsible is the best safety guard. His stance rejects a lobbying push that gained urgency after Anthropic’s Mythos highlighted AI-enabled cyber threats.
Four paying subscribers have sued OpenAI, Anthropic, Google, and SpaceXAI under Section 1 of the Sherman Act, alleging that the CEOs' public statements and coordinated efforts to slow AI development unlawfully restrained competition. The case tests whether safety-motivated coordination among rivals crosses antitrust lines.
In today's Smoke evaluation, Grok 4's material constraint score dropped from 100.00 to 84.20, a 15.8-point decline that pulled its main leaderboard score from 95.44 down to 90.86. Code execution rose, but engineering judgment also fell sharply, pointing to volatility in material-sensitive tasks.
In today's Smoke evaluation, Claude Sonnet 4.6's material constraint score dropped from 100.00 to 84.20, while code execution rose from 50.00 to 97.00, pushing its main leaderboard score from 72.50 to 91.24. The opposite swings likely reflect small-sample randomness rather than a real capability decline.
On 2026-09-22, the YZ Index Smoke quick test covered 10 models, with GPT-o3 ranking first at 96.01. The Smoke test is a daily 10-question quick assessment useful for observing short-term signals, not equivalent to full weekly rankings.
On September 18, 2026, OpenAI, Anthropic, and xAI signed the AEF-1 embedded evaluation standard, while more than 100 researchers including Geoffrey Hinton and Stuart Russell signed an open letter warning that the existing framework is insufficient to prevent conflicts of interest. The simultaneous moves highlight the tension between principles and enforcement.
Anthropic released a prototype R&D automation index showing that by August 2026, Claude was handling 26% of the company's AI research and engineering work at the AL4 level, up from under 1% in February. The framework, built on a frozen task tree of 542 nodes, also details oversight data and compute allocation for safety work.
The U.S. has formally proposed a bilateral AI national security incident notification mechanism, with both sides agreeing to establish a dedicated working group less than 96 hours before the Trump-Xi summit. The talks also covered trade committee arrangements and possible tariff cuts on non-sensitive goods.
A Financial Times investigation found that Alphabet, Meta, Nvidia, Broadcom, Oracle, Amazon and others have moved more than $300 billion in AI infrastructure debt off their balance sheets through residual value guarantees and special purpose vehicles. The structures comply with GAAP, but they create hidden systemic exposure as 25-year debt funds assets with useful lives of only 5 to 7 years.
This week, 454 translation tasks were completed by 4 models. A blind multi-model comparison of 3 sampled articles found claude-sonnet-4.6 to be the best overall, with an average score of 9/10.
Anthropic is weighing whether to release a new frontier AI model before its planned IPO, balancing investor confidence, internal safety assessments, and competition with OpenAI’s GPT-6 Astra. The decision comes as its IPO timeline slips and its valuation is discussed in the trillion-dollar range.
In a CBS News interview, NVIDIA CEO Jensen Huang said the chance of an AI apocalypse by 2030 is 0%, even as Anthropic’s Dario Amodei, OpenAI’s Sam Altman, and others at AI labs he supplies are publicly urging the industry to slow down. The article examines the conflict between Huang’s business incentives and the warnings from those building frontier AI systems.
In today's Smoke evaluation, Claude Sonnet 4.6's main leaderboard score dropped 17 points, driven by code execution falling from full marks to 50, a swing most likely explained by the small daily question sample rather than a genuine decline in model capability.
In today's Smoke evaluation, GPT-5.5's main leaderboard score fell from 86.50 to 76.44, a 10.1-point decline. The drop was driven entirely by a 25-point decrease in the Code Execution dimension, while Material Constraints rose 8.2 points.
The 2026-09-21 YZ Index Smoke quick test covers 10 models, with Grok 4 leading the day at 95.44 points. Smoke is a daily 10-question quick test, useful for observing short-term signals and not equivalent to Full weekly leaderboard conclusions.
