- Published on
DeepSWE Benchmark Crushes Claude Fable 5, Clues Point to Zhipu
- Authors
- Name
- DeepHorus
On August 20th, OpenRouter, the world's largest AI aggregation platform, quietly launched a mysterious AI model named 'Ox Alpha' and announced it would be free for a week. This move quickly stirred waves in the developer community. What's even more striking is that independent developer Ben Davis used this model to achieve a remarkable score of about 80% on the DeepSWE benchmark, far surpassing Anthropic's Claude Fable 5 (65%) and OpenAI's GPT-5.6 Sol (52%), both of which are currently recognized as top-tier coding models in the industry. Davis himself was puzzled by the anonymous model's outstanding performance, but its formidable capabilities have already sparked widespread speculation.
The background of this event traces back to the fierce competition in the current AI large model landscape. Major companies are continuously releasing models with stronger performance, and DeepSWE, as an authoritative benchmark that measures an AI's ability to read real codebases, identify vulnerabilities, and generate effective patches, is a key indicator of model performance in complex software engineering tasks. The appearance of Ox Alpha coincides with the window period when major companies are about to release new models, and its anonymous and free strategy seems more like a meticulously planned 'technical debut.'
According to the technical specifications disclosed by OpenRouter, Ox Alpha's parameter strength is not to be underestimated. It supports a context window of up to 1,048,576 tokens, can process multimodal inputs including text, images, and videos, and has a maximum output length per response of 131,072 tokens. These specifications indicate that it is not a lightweight demo model but a heavyweight contender capable of handling complex, long-sequence tasks. Meanwhile, the terminal programming agent OpenCode also offers this model through its free tier, claiming it can process up to 100 trillion tokens daily, further confirming its massive underlying computational support.
Despite the model's mysterious origins, the tech community's analysis has gradually unveiled its veil. Researchers found through tokenizer fingerprint analysis of 25 prompts that Ox Alpha's tokenization count almost exactly matches that of Zhipu AI's GLM-5.3 model, with only a constant 75-token packaging layer deviation. Additionally, the model's video encoder behavior patterns and API response characteristics are highly consistent with Zhipu's architecture. These clues strongly suggest that Ox Alpha is likely Zhipu's upcoming new-generation large model.
This event has had multiple impacts on the industry. First, it directly challenges existing model performance rankings, demonstrating that domestic large models may have reached or even surpassed international leading levels in core coding capabilities. Second, for developers, the one-week free window provides a zero-cost opportunity to experience top-tier models, which could attract a large number of users for testing and feedback, accumulating valuable initial data for Zhipu. Furthermore, the anonymous release strategy has created buzz and suspense, generating massive market attention for the eventual brand reveal.
The industry's reactions are mixed. Some analysts view this as Zhipu's clever market warm-up activity, aimed at building recognition through hard power rather than brand effect. Another AI startup CTO commented: 'If the DeepSWE scores are genuine, this is indeed impressive. But we need to see more evaluations on diverse tasks.' Meanwhile, developers and users have shown keen interest, with discussions on social media about 'who exactly is Ox Alpha' becoming a hot topic, and many are excited about the sudden progress of domestic models.
Looking ahead, the most direct expectation is that Zhipu will likely officially announce this model and reveal its true identity and commercial plans after Ox Alpha's free period ends. The user feedback and performance data accumulated during this 'anonymous test' will serve as an important reference for its subsequent optimization and market strategy. If the model proves to be powerful, it could quickly enter the commercial API service market, competing directly with GPT, Claude, and others.
From a macro perspective, the emergence of Ox Alpha is a microcosm of the intensifying global AI race. Chinese AI companies are transitioning from followers to parallel players and evenć±éš leaders. Zhipu's choice to launch a 'surprise attack' in the coding capability field, which is crucial to the developer ecosystem, demonstrates its differentiated competition strategy. This event also foreshadows that future large model releases may become more diverse, with anonymous testing that 'speaks through products and performance' potentially becoming a new competitive tactic, accelerating changes in the industry landscape.
