Two major announcements on April 23, 2026, underscore the rapidly escalating race in artificial intelligence, particularly in agentic capabilities and coding performance. Chinese tech giant Tencent released its Hy3 preview model, while OpenAI launched GPT-5.5, both targeting developers and power users with significantly improved benchmarks.
Tencent Hy3 Preview: Efficiency-Focused Open-Source Model
Tencent quietly introduced its Hy3 preview model, described as the company's most capable AI yet and built after a full infrastructure rebuild. The model is a 295 billion parameter Mixture-of-Experts (MoE) architecture, with only 21 billion active parameters per query. This design allows for substantial cost savings while maintaining high output quality. Tencent explicitly argues that 295 billion parameters represent the optimal sweet spot where reasoning fully matures but the cost of adding more parameters stops paying off, a step back from their previous Hy2 model which had over 400 billion parameters.
The performance gains are dramatic. On SWE-bench Verified, a coding benchmark testing real GitHub bug fixes, Hy3 jumped from Hy2's 53.0% to 74.4%—a 40% improvement. On Terminal-Bench 2.0, which measures autonomous execution in a real command-line environment, it leaped from 23.2% to 54.4%. On BrowseComp, testing complex web research tasks, it achieved 67.1%, up from 28.7%. In reasoning, Hy3 topped all Chinese competitors on Tsinghua University's math PhD qualifying exam with an 88.4 avg@3 score and scored 87.8 on the CHSBO 2025 biology olympiad, the highest among Chinese models.
The model is already live across Tencent's app ecosystem including Yuanbao, QQ, Tencent Docs, and developer tools CodeBuddy and WorkBuddy, where first-token latency dropped 54% and end-to-end generation time fell 47%. API access on Tencent Cloud starts at approximately $0.18 per million input tokens and $0.59 per million output tokens, with personal token plans from about $4.10 per month. Hy3 was open-sourced on GitHub, Hugging Face, and ModelScope, and began training in late January 2026—a remarkably fast three-month development cycle attributed to a February infrastructure overhaul led by chief AI scientist Yao Shunyu.
OpenAI GPT-5.5: Smarter, Faster, and Pricier
OpenAI released GPT-5.5 on the same day, positioning it as a model designed for agentic computer use and complex multi-step tasks. The model is rolling out to Plus, Pro, Business, and Enterprise subscribers on ChatGPT and Codex, but not to free users. On Terminal-Bench 2.0, GPT-5.5 scored 82.7%, significantly beating Claude Opus 4.7 at 69.4% and Gemini 3.1 Pro at 68.5%. On GDPval, which tests knowledge work across 44 real occupations, it matches or beats industry professionals in 84.9% of comparisons. On SWE-Bench Pro, it reached 58.6%, though Claude Opus 4.7 scored 64.3% (OpenAI claims potential memorization in Anthropic's results). The high-end GPT-5.5 Pro scored 90.1% on BrowseComp, ahead of Gemini 3.1 Pro's 85.9%.
Pricing, however, has increased significantly. The API will charge $5 per million input tokens and $30 per million output tokens, compared to GPT-5.4's $2.50/$15.00. GPT-5.5 Pro API pricing remains at $30/$180 per million tokens. OpenAI CEO Sam Altman argued on X that token efficiency gains offset the cost, as GPT-5.5 completes the same Codex tasks with fewer tokens. The launch comes just seven weeks after GPT-5.4, reflecting the accelerated release tempo in the AI industry.
Both announcements highlight the growing emphasis on agentic AI—models that can autonomously execute complex, multi-step tasks—and the intense competition between US and Chinese AI labs to deliver powerful, cost-efficient models for developers and enterprises.
