Quick Take

This comparison evaluates Tencent’s Hy3-preview (released April 2026) against OpenAI’s GPT-5.3 Codex (xhigh) (released February 2026). Hy3-preview positions itself as a high-velocity, budget-friendly model, whereas GPT-5.3 Codex (xhigh) focuses on maximizing intelligence and benchmark scores at a premium price.

Benchmark Read

GPT-5.3 Codex (xhigh) consistently outperforms Hy3-preview across most standardized metrics. It achieves an intelligence index of 44.3 compared to Hy3-preview’s 33.6. In specific benchmarks, GPT-5.3 Codex (xhigh) leads with a GPQA score of 0.915 (vs. 0.867), HLE of 0.399 (vs. 0.255), and SciCode of 0.532 (vs. 0.412). However, Hy3-preview demonstrates a notable advantage in the TAU2 benchmark, scoring 0.927 against GPT-5.3 Codex’s 0.860, suggesting specific strengths in that domain.

Cost and Speed

There is a stark contrast in operational efficiency. Hy3-preview is significantly more affordable, with a blended cost of $0.20/1M tokens, compared to the $4.81/1M blended cost of GPT-5.3 Codex (xhigh).

Performance metrics further highlight these differences. Hy3-preview delivers an output speed of 153.661 tok/s with a rapid time-to-first-token (TTFT) of 1.633s. In contrast, GPT-5.3 Codex (xhigh) operates at 105.49 tok/s and exhibits a substantial TTFT of 52.556s, which may impact real-time user experience.

Best Fit

Hy3-preview is best suited for developers and enterprises requiring high-throughput, low-latency reasoning at scale without the premium cost associated with frontier-level models. GPT-5.3 Codex (xhigh) is the preferred choice for complex, high-stakes tasks where maximum intelligence and accuracy are non-negotiable, and where the higher latency and cost are justified by performance gains.