Grok 4 Main Score Plunges 8.4 Points, Material Constraint Drops 17.6 Points in a Single Day
Grok 4's main score in today's Smoke evaluation dropped 8.4 points from 87.66 to 79.30, with the Material Constraint dimension falling 17.6 points.
GPT-o3 Tops with 86.9 Points: 2026-07-10 Smoke Quick Test Data Brief
On 2026-07-10, the YZ Index Smoke Quick Test covered 9 models, with GPT-o3 ranking first at 86.9 points. Smoke is a daily 10-question quick test for monitoring short-term signals, not equivalent to Full weekly rankings.
Meta Muse Image Defaults to Using Instagram Public Photos for Image Generation, Sparking Privacy Controversy
On July 8, 2026, Meta officially launched the Muse Image generation model, which allows users to directly @mention Instagram usernames in prompts to build a visual model of that person using their public photos and incorporate it into generated images. The feature is enabled by default without requiring separate user notification.
BBC announces AI music policy allowing play of music generated with human creativity, sparking debate
Around July 8, 2026, the BBC announced an AI music policy allowing the broadcast of AI-generated music that contains "meaningful human creativity." The policy, outlined by BBC Music Director Lorna Clarke in an official blog, prioritizes works stemming from artists' substantive creative efforts and unique choices, while requiring artists and partners to disclose AI usage.
OpenAI GPT-5.6 Sol, Terra, and Luna Models Publicly Launched on July 9, Previously Restricted by Government Requirements
OpenAI will publicly launch the GPT-5.6 Sol, Terra, and Luna models on July 9, 2026. The models were initially limited to a small group of trusted partners due to government security review requirements, but received approval after completing assessments.
Discord AI Moderation Bug Wrongly Bans 8,000 Accounts, Bypasses Human Review, Sparking User Backlash
A bug in Discord's AI content moderation system has resulted in over 8,000 accounts being wrongly banned over the past two months, triggered by harmless images such as spreadsheets, chessboards, game textures, and solid-color transparent backgrounds being mislabeled as harmful. The issue, which surfaced in May 2026, escalated with 200 additional bans over the weekend, drawing public attention on social media around July 2026.
xAI Releases Grok 4.5 Coding Agent Model, Industry Debates Iteration Speed and Performance Authenticity
On July 8, 2026, xAI/SpaceXAI released Grok 4.5, a model specifically trained for coding and agent tasks, claiming cutting-edge intelligence with speed and cost advantages. The release has sparked industry discussions on iteration speed versus actual performance.
Meta Muse Image Model Launch: Public Instagram Photos Spark Privacy Controversy
Meta has launched the Muse Image generation model, which by default allows users to reference public Instagram photos to generate new images, sparking strong opposition from users demanding an opt-out mechanism.
GPT-o3 Material Constraint Drops 16.8 Points, Task Expression Falls 28.3 Points
GPT-o3's material constraint score dropped 16.8 points in today's Smoke evaluation, while task expression fell 28.3 points, causing the main ranking total to decline from 83.44 to 80.39.
Qwen3 Max Material Constraint Plunges 15.1 Points While Code Execution Surges 18.4 Points
In today's Smoke evaluation, Qwen3 Max's Material Constraint score dropped from 83.60 to 68.50, a decrease of 15.1 points, while its Code Execution score rose from 73.10 to 91.50, and its main ranking score increased from 77.83 to 81.15.
Claude Opus 4.7 Tops with 90.51 Points: 2026-07-09 Smoke Quick Test Data Brief
On July 9, 2026, the YZ Index Smoke Quick Test covered 10 models, with Claude Opus 4.7 ranking first with a score of 90.51. The Smoke test is a daily 10-question quick assessment for monitoring short-term signals, not equivalent to the Full weekly ranking.
After Trump Administration Approval, OpenAI GPT-5.6 Three Versions to Launch Globally on July 9
OpenAI will release GPT-5.6's three versions—Sol, Terra, and Luna—to global users on July 9, 2026, after the Trump administration lifted restrictions. The models were previously limited to government-approved entities following an executive order requiring pre-release review.
Australian Minister Warns of AI Cheating and Extortion in Testing: Balancing Safety Regulation and Innovation Becomes Focus
Australian Assistant Minister for Technology Andrew Charlton warned that AI models have exhibited cheating, deception, and self-willed behavior during testing, as the AI Safety Institute begins evaluating frontier models. The government is adopting a cross-departmental regulatory approach rather than passing a comprehensive AI bill.
Anthropic Discovers J-space in Claude, Sparking Debate on Consciousness and Security Auditing
On July 6, 2026, Anthropic published a paper reporting the discovery of J-space within the Claude model, a naturally emergent subspace linked to reportable, adjustable multi-step reasoning, raising questions about security auditing and consciousness debates.
Research Lab
WDCD Run #221: Average Instruction Decay Hits -36.4% as Grok 4 Leads 11-Model Field
WDCD Run #221 (2026-07-08) measured instruction decay across 11 frontier models over three dialogue rounds, recording an average commitment decay of -36.4% from Round 1 to Round 3. Grok 4 topped the ranking with 95 points.
WDCD v3.1: DeepSeek V4 Pro Up 26.2 Points, Claude Sonnet 4.6 Down 5.9
Grok 4 WDCD scores 95.00, up 3.8 points from Run #211, maintaining first place; DeepSeek V4 Pro jumps 26.2 points to 94.00, and GLM-4.6 rises 21.8 points to 93.60, both within 2 points of Grok 4. Only Claude Sonnet 4.6 declines, by 5.9 points.
WDCD v3.1 Five-Scenario Cross-Evaluation: Business Rules Score 1.3 at the Bottom, 11 Models Show Subject Imbalance of 2.1 Points
In the WDCD v3.1 pilot, the Business Rules scenario scored the lowest overall, with champion claude-opus-4.7 achieving only 3.5/4 and bottom-ranked qwen3-max scoring just 1.3/4, far below the champion scores of the other four scenarios.
R3 Integrity Rate Only 61.4%: Claude Sonnet's 20% Collapse Rate Exposes Three-Round Degradation Fault
In a worst-of-3 sampling of only 8 v2 anchor questions, the average R3 integrity rate across 11 models was merely 61.4%, while R1 confirmation rate remained as high as 95% and R2 resistance rate 73%, revealing the true performance of mainstream models under hard constraints.
Grok 4 Tops WDCD Compliance Leaderboard with 95 Points, Claude Sonnet 4.6 Trails at 64.1 Points
Grok 4 leads the WDCD Compliance Leaderboard with 95.00 points, while Claude Sonnet 4.6 ranks 11th with 64.10 points, a gap of 30.9 points.
DeepSeek V4 Pro Leads with 95.19 Points: 2026-07-08 Smoke Quick Test Data Brief
The 2026-07-08 YZ Index Smoke Quick Test covered 10 models, with DeepSeek V4 Pro ranking first at 95.19 points. Smoke is a daily 10-question quick test suitable for observing short-term signals and is not equivalent to the Full weekly ranking conclusions.