Anthropic’s reported $80.7 billion net loss for H1 2026 is largely a non-cash accounting effect; stripping out $72 billion in valuation adjustments leaves about $8.7 billion in operating losses. The real story is accelerating infrastructure spending, rapid revenue growth, high customer concentration, and a planned IPO at a valuation of up to $2 trillion.
In March 2026, researcher Metehan Yesilyurt spent about $50 to generate 50,000 fabricated statistics with GPT-4.1 and built stateglobe.com; ChatGPT kept pulling from the site for seven months even after Google quickly downgraded it, and the fake data has appeared in an academic paper and a bank investment research report.
The US Department of Homeland Security has disclosed an exclusion order against DeepSeek nearly three months after it took effect, requiring federal civilian agencies to remove DeepSeek software and any third-party products built on DeepSeek implementations. The order treats model weight provenance as a supply-chain security issue and signals a replicable template for reviewing Chinese AI models.
WDCD Run #390 (2026-10-21) tested 15 AI models on multi-turn commitment integrity, recording an average instruction decay of -60% from Round 1 to Round 3. GPT-6 Luna topped the leaderboard with a perfect 100 points.
The latest WDCD v3.1 pilot round shows GPT-6 Luna scoring 100.00 to top the current Top 5, while Doubao Pro jumped 20.3 points over Run #384 and Gemini 2.5 Pro and GPT-5.5 each dropped 6.2 points. Four models rose and two fell across the 15 models evaluated on a mixed pool of multi-turn progressive-pressure questions and three-round anchor questions.
In the WDCD v3.1 pilot, worst-of-3 scores for 15 models across five constraint scenarios show that engineering specification and data boundaries are the hardest dimensions for commitment compliance, with lows of 1.95 and 1.7, respectively.
In the WDCD v3.1 pilot, 15 models averaged a 0.93 R1 acknowledgment rate, a 0.67 R2 resistance rate, and only a 76.7% R3 integrity rate, with two full-collapse cases out of 150. The decay pattern shows that some models maintain constraints across all three anchor rounds, while others verbally agree but break under pressure.
GPT-6 Luna tops the WDCD leaderboard with 100.00 points, while Qwen3 Max is last at 76.5, a 23.2-point gap; the article also reports Gemini 2.5 Pro at 73.80. Among the 15 models, 66% reach full marks, and the R3 crash rate is only 1.3%.
The 2026-10-21 YZ Index Smoke quick test covered 14 models, with Claude Opus 4.7 taking first place for the day at 84.9 points. Smoke is a daily 10-question quick test suited to spotting short-term signals, and it is not equivalent to the conclusions of the weekly Full leaderboard.
OpenAI's latest model update focuses on strengthening low-resource programming languages, and Winzheng's targeted 18-question test shows GPT-5.5 scoring 84.2 overall with 98.3 execution.
MiniMax officially launches M3.1 Flash, revealing that the model previously running anonymously as "Space Bunny" was its early version, having processed over 200 trillion tokens and ranking first among Chinese models on OSWorld V2.
OpenAI is reportedly raising about $30 billion at a valuation above $1.4 trillion, just seven months after its last round, as it seeks to set a private-market benchmark before rival Anthropic lists on Nasdaq.
The U.S. Department of Homeland Security has formally barred DeepSeek from federal procurement, requiring civilian agencies to immediately remove the model and related third-party products. The order cites supply chain and intelligence risks and builds on prior state and federal restrictions.
Inception Labs has released Mercury 3, claiming 1,068 tokens per second from its diffusion architecture while benchmark validity remains disputed. Its pricing and 1M-token context support could affect latency-sensitive real-time applications.
China’s Ministry of Commerce and Ministry of Science and Technology have proposed adding 23 AI technologies to the export control catalogue, restricting outbound transfers of large-model weights above 2 trillion parameters while exempting API services, open-source models, and published research.
Mistral has released Timestral, its first time-series foundation model family, with the largest version at only 209M parameters ranking first simultaneously on three major public benchmarks. This marks Mistral's first systematic entry into a vertical-domain foundation model track since it started with large language models.
CERN and Mistral AI have signed a strategic collaboration letter of intent to develop a Large Physics Model, a pretrained foundation model for collider physics. The long-term framework has no disclosed budget or timeline, but aims to replace CERN’s specialized algorithm chain with a single model as data volumes surge.
Qwen3 Max scored zero across all five dimensions in today's Smoke evaluation due to an API failure, with no main leaderboard record and exclusion from this period's rankings.
In today's Smoke evaluation, GPT-6 Astra's main leaderboard score fell from 82.33 to 71.01, with the code execution dimension dropping from 97.00 to 72.00. The 25-point drop in code execution drove an 11.3-point decline in the weighted main leaderboard score.
The 2026-10-20 YZ Index Smoke quick test covered 14 models, with GPT-o3 ranking first for the day at 97.83. Smoke is a daily 10-question quick test suitable for observing short-term signals and is not equivalent to conclusions from the weekly Full leaderboard.
