On July 7, 2026, Meta launched the Muse Image feature, allowing users to generate images by @-mentioning public Instagram accounts, but withdrew it three days later following backlash over opt-out defaults.
SK Hynix completed a $26.5 billion IPO on the Nasdaq, the largest by a foreign company in U.S. capital markets, with proceeds mainly used to expand HBM production lines for AI servers.
Meta's newly launched AI image generator Muse faces user and regulatory scrutiny due to its default data usage settings that automatically use public Instagram photos for model training, leading to potential takedown pressure.
DeepSeek V4 Pro's main ranking score in today's Smoke evaluation plummeted from 96.99 to 80.10, a drop of 16.9 points. Code execution alone fell 28 points, contributing the vast majority of the decline.
In the Smoke evaluation, Claude Opus 4.7's main leaderboard score dropped from 96.99 to 82.95, and its Code Execution dimension fell from 100.00 to 69.00, a decline of 31 points.
The YZ Index Smoke Quick Test on July 13, 2026 covered 10 models, with Gemini 2.5 Pro scoring 100 to top the daily leaderboard. This daily 10-question test is designed for short-term signal observation and does not replace the Full weekly ranking conclusion.
Meta Platforms officially released the Muse Spark 1.1 model on July 9, 2026, opening a public preview to US developers with a $20 free credit and usage-based pricing starting at $1.25 per million input tokens.
Meta launched Muse Image in early July 2026 but quickly withdrew it after privacy backlash. The feature automatically used public Instagram content, sparking opposition from creators and SAG-AFTRA.
OpenAI and Google have confirmed providing advanced AI model access to subsidiaries of Alibaba, Baidu, and Tencent in Singapore that are on the U.S. Department of Defense's Section 1260H blacklist, which identifies entities linked to the People's Liberation Army. The event highlights tensions between commercial expansion and national security.
Apple has filed a lawsuit against OpenAI in the U.S. District Court for the Northern District of California, accusing the company of illegally obtaining hardware-related trade secrets through the recruitment of former Apple employees. The lawsuit also names OpenAI's Chief Hardware Officer Tang Tan and former Apple engineer Chang Liu as key figures, alleging that OpenAI has contacted at least two Apple suppliers to extract process details.
In July 2026, Meta launched the Muse Image feature on Instagram, allowing users to mention public accounts to have Meta AI generate new images using their public photos as references. After several days of strong opposition, the feature was withdrawn.
On July 11, 2026, Apple filed a lawsuit in the U.S. District Court for the Northern District of California, accusing OpenAI, its hardware head Tang Tan, former Apple engineer Chang Liu, and io Products (acquired by OpenAI) of systematically stealing Apple's AI hardware design, manufacturing processes, supplier collaboration details, and other trade secrets to advance its own consumer hardware projects.
WDCD Run #227 (2026-07-12) evaluated 11 frontier models on multi-turn commitment integrity, with Grok 4 and DeepSeek V4 Pro tying at 91.4 points and average instruction decay measured at -2.8% between Round 1 and Round 3.
In the latest WDCD v3.1 compliance test, all nine evaluated models showed positive score changes, with no model declining. DeepSeek V4 Pro surged 23.6 points to 91.36, just 0.04 points behind Grok 4's 91.40.
In the WDCD v3.1 five-scenario review, the safety compliance scenario had the lowest overall scores, with Qwen3-Max scoring only 2.13/4 and Grok-4 reaching 3.86/4, a gap of 1.73 points.
In the WDCD v3.1 pilot, 11 models tested under three-round anchoring showed an average R3 integrity rate of only 50.7% with 28 zero-score cases, revealing severe promise-keeping degradation from R1 confirmation to R3 collapse.
Grok 4 tops the WDCD v3.1 commitment test with 91.40 points, while Qwen3 Max scores the lowest at 64.88 points, showing a significant gap between top and bottom models.
In today's Smoke evaluation, GLM-4.6 experienced a sharp 25-point drop in material constraint and a 50-point surge in code execution, driving the leaderboard score up by 16.3 points. The fluctuations are more likely due to question sampling variance than model degradation, though further monitoring is needed.
Smoke 7-day evaluation data from 2026-07-08 to 2026-07-12 shows Gemini 2.5 Pro leading with a 34.3-point uptrend, while GLM-4.6 records the highest volatility of 40.5 points.
On July 12, 2026, the YZ Index Smoke Quick Test covered 11 models, with Claude Opus 4.7, Claude Sonnet 4.6, DeepSeek V4 Pro, and Grok 4 all scoring 96.99 to tie for first place. The Smoke test uses 10 daily questions, serving as a short-term signal rather than a definitive weekly ranking conclusion.
