YZ Index Major Overhaul: 7 New Models Including GPT-5.5, Claude Opus 4.7, and DeepSeek V4 Launch Simultaneously as 9 Veterans Retire

On May 1, 2026, YZ Index completed its largest evaluation roster update since launching last year—replacing 9 models and introducing 7 new flagships at once, with the evaluation system undergoing a complete generational overhaul.

This is not a routine version bump. Over the past month, seven service providers—OpenAI, Anthropic, DeepSeek, Google, xAI, Alibaba, and Baidu—have nearly simultaneously completed generational leaps. The old models have either stopped receiving updates, or their APIs have been redirected to new versions. As an evaluation system that tracks the real-world performance of AI models on a weekly basis, YZ Index cannot keep running last year's contestants on this year's track.

Who's Out, Who's In

First, the biggest change: GPT-4o exits the YZ Index evaluation.

GPT-4o has been one of the benchmark models for YZ Index since last year. But over the past year, OpenAI has successively released GPT-5, 5.1, 5.2, 5.3, 5.4, and most recently GPT-5.5 (released April 23, 2026). Continuing to evaluate GPT-4o no longer offers any reference value—it would be like comparing photo quality between an iPhone 12 and an iPhone 16's contemporaries.

Similar situations have unfolded with every service provider:

• Anthropic: Claude Opus 4.6 → Claude Opus 4.7. Sonnet 4.6 is retained as the latest version on the value-for-money tier.
• DeepSeek: V3 and R1 have been merged into the V4 architecture. The old deepseek-chat and deepseek-reasoner APIs have been redirected to deepseek-v4-flash. Evaluation now uses the new DeepSeek V4 Pro.
• Google: Gemini 2.5 Pro is retained, with the addition of Gemini 3.1 Pro (preview). This marks the first time YZ Index evaluates two generations of models from the same provider simultaneously—we want to see the real performance level of 3.1 in its preview stage.
• xAI: Grok 3 → Grok 4.
• Alibaba: Qwen Max → Qwen3 Max. Tongyi Qianwen has now reached its third generation.
• Baidu: ERNIE Bot 4.0 → ERNIE Bot 4.5.
• ByteDance: 豆包 Pro remains unchanged.

Why Replace So Many at Once

Because the update cadence in the AI industry is no longer measured in years, but in months.

OpenAI has released 6 major version numbers (GPT-5 through 5.5) in the past 12 months. Claude has iterated from Opus 4 to 4.7. DeepSeek has even outright deprecated its old architecture. If YZ Index continued evaluating these old models that no longer represent each provider's strongest capabilities, the leaderboard would become a historical museum—telling you who was strong six months ago, but offering no help for today's selection decisions.

Our principle is simple: Every model on the leaderboard should be one you can use today and worth using.

How the New Lineup Will Affect Rankings

To be frank: The first round of data will be relatively raw.

YZ Index uses a 5-run rolling average to smooth out single-instance fluctuations. New models have no historical data, so first-time evaluation rankings are based entirely on a single performance, which may be skewed high or low. Rankings will only reflect true, stable performance after running for 5 consecutive weeks.

This is exactly why we're highlighting it in this update—if you see a new model ranked unusually high or unusually low, don't rush to conclusions. Give it 5 weeks.

Daily Services Upgraded in Sync

This update affects more than just the evaluation leaderboard. All AI-driven features at Winzheng—news signal discovery, fact-checking, article generation, translation, social media analysis—have all been simultaneously switched to the latest models:

• Signal Discovery: Grok 4 (real-time X platform search)
• Fact-Checking: Gemini 2.5 Pro (Google Grounding verification)
• Article Writing: Claude Sonnet 4.6 (best writing quality)
• Low-Cost Batch Tasks: DeepSeek V4 Flash (cost under $0.5/million tokens)

All task configurations are centrally managed through a unified AI Management Center (three-tier architecture: provider → model → task). Future model upgrades will only require changing a single field in the backend, and all site-wide features will automatically follow—no more file-by-file code changes.

What About Historical Data

All preserved. Historical evaluation data, weekly change records, and incident reports for retired models remain intact in the database and can be viewed in the historical records at any time. YZ Index does not delete any evaluation data—this is our commitment to data integrity.

The new lineup will make its debut in the next weekly evaluation. At that point, you'll see a fresh batch of names appearing on the leaderboard's homepage.

Let's see which of these 2026 newcomers truly deserves the title of all-around champion.