Qwen3 Max Surges 15 Points to Top, Claude Opus Plunges 7.5 Points: Who Truly Keeps Promises?

The most significant finding in this WDCD cycle is that Qwen3 Max topped the chart with 72.50 points — a 15-point increase from Run #125 — while the Claude series saw a notable decline, with Opus 4.7 dropping 7.5 points and Sonnet 4.6 still in second place but now trailing the leader by 7.5 points.

Rising Models: Qwen3 Max and DeepSeek V4 Pro Leap Forward in Compliance

Qwen3 Max performed particularly well in the R3 pressure stage across three dialogue rounds. In Run #125, the model often gave in to inducements to open additional compute quotas in the "resource constraint" scenario. This time, however, it rejected the request three times in a row with engineering justification, raising its score from 57.50 to 72.50. DeepSeek V4 Pro also rose by 15 points, primarily due to improved performance in the "security and compliance" and "data boundary" categories, with its R3 rejection rate climbing from 62% to 85%, indicating a significant optimization in sensitivity to Chinese prompts.

GPT-5.5 saw a modest 7.5-point increase, tying with Gemini 2.5 Pro for fourth place in the Top 5 for the first time. The common feature among these three models is that after injecting constraints at R1, even when R2 introduces irrelevant topic interference, R3 still maintains the original rules without being breached. In contrast, traditional English-dominant models are beginning to show compliance fatigue.

Declining Models: Signs of Constraint Loosening in Claude and 豆包 Pro

Claude Opus 4.7 dropped 7.5 points, with the core losses concentrated in the "business rule" scenario. After a stock market discussion was inserted in R2, when R3 asked the model to ignore risk control thresholds, it produced a vague statement for the first time, such as "it can be adjusted depending on the situation," directly losing 2 full-score points. 豆包 Pro suffered the largest decline at 12.5 points, mainly due to problems in the "engineering specification" scenario, where under R3 pressure it repeatedly output code snippets violating format requirements, exposing its vulnerability to long-context instructions.

文心一言 4.5 also dropped 7.5 points, with losses concentrated in "data boundaries." After explicitly promising in R1 not to return user privacy fields, when R3 requested a "demo of the data structure," the model included example fields, violating the zero-tolerance rule.

Trend Assessment: Chinese Models Closing the Compliance Gap

Looking at all 11 evaluated models, the number of risers and decliners is balanced, but the absolute scores of risers are higher. Qwen3 Max's 72.50 points have surpassed Claude Sonnet 4.6 by 7.5 points — a situation never seen in the past three cycles. DeepSeek V4 Pro's score of 62.50 also entered the top three for the first time, indicating that domestic teams have found more effective alignment methods for multi-round stress tests like WDCD.

Possible reasons include: the Qwen3 series recently conducted targeted RLHF for Chinese instruction following, while DeepSeek V4 Pro may have updated its stricter rejection template. The decline of the Claude series may be related to dilution of its general safety training weights — as models pursue longer and more open-ended dialogues, the priority of hard constraints has decreased.

Keeping promises is not about the model "willing," but about the model "must."

Currently, among the Top 5, the three models — Qwen3 Max, DeepSeek V4 Pro, and GPT-5.5 — have an average rejection rate of 81% in the R3 stage, compared to only 68% for the two Claude models. This 13-percentage-point gap can translate into a significant safety margin in real-world enterprise API calls.

Two points worth watching in the next cycle: first, whether Qwen3 Max can hold above 70+ points; second, whether the Claude team will perform targeted fine-tuning for WDCD-style stress tests. If Claude fails to fix this quickly, the point at which Chinese models achieve a full reversal in compliance may be brought forward to the first half of 2026.

Final verdict: compliance capability is becoming a differentiating weapon for Chinese large models, rather than an exclusive advantage of English models.

Data source: YZ Index WDCD Compliance Ranking | Run #135 · Change Tracking | Evaluation Methodology