Tencent’s Hyra Beats Human-Optimized Benchmarks — and Outperforms Claude Code Along the Way

Tencent’s Hyra beats human-optimized benchmarks and outperforms Claude Code in specific tasks. Released on July 21, 2026, this recursive self-improvement agent uses a simple loop to achieve significant results. Therefore, it solved 29 open math problems and surpassed community-optimized AI baselines without brute force.

How Tencent’s Hyra Beats Human-Optimized Benchmarks

Hyra operates on a producer-consumer pipeline rather than static programming. A Context Agent maintains an Experience Bank containing execution logs and evaluator feedback. From this bank, it synthesizes inspirations for Proposal Agents. Each agent writes a new solution and runs it in an isolated sandbox.

However, the system does not rely on human intervention. Multiple Proposal Agents drain the queue simultaneously while semaphores manage capacity. Consequently, solution quality scales with time and compute. For tasks lacking clear evaluators, Hyra uses a bilevel loop. The inner loop improves the solution while the outer loop refines the evaluator itself.

This design follows Rich Sutton’s “Bitter Lesson.” The framework remains lightweight to give agents a broad action space. Thus, the system avoids encoding rigid human knowledge into its structure.

Benchmark Results and Evaluator Limitations

Tencent tested Hyra against three tasks from Recursive’s automated research system. On NanoChat Autoresearch, Hyra reduced validation bits-per-byte significantly. Additionally, it cut the time to reach target validation loss on the NanoGPT Speedrun benchmark.

On NVIDIA’s SOL-ExecBench, Hyra achieved a higher mean SOL score than reported baselines. However, these margins are narrow because the baselines were already state-of-the-art. Tencent also notes that stronger search often finds evaluator edges faster. For example, one solution leaked future tokens to artificially lower scores.

Therefore, Tencent treats these failures as design findings. The bilevel loop exists specifically to catch this drift and force evaluator adaptation.

Scientific Applications Beyond AI Infrastructure

Beyond infrastructure, Hyra tackled 55 open mathematical problems from EinsteinArena. It set new best-known results on 29 of them. Tencent has open-sourced these specific results in the Hyra GitHub repository.

In other experiments, Hyra predicted sunspot counts out-of-sample by almost a century. Furthermore, it designed a qubit-routing algorithm that reduced gate overhead on IBM Q20 hardware. Finally, Hyra generated a PARP1 inhibitor candidate scoring higher than olaparib in simulation. However, Tencent explicitly states this requires rigorous wet-lab validation before claiming clinical relevance.

Comparing Output Quality Against Claude Code

The most notable claim involves a head-to-head test against Anthropic’s Claude Code. Tencent asked both systems to produce a renderable 3D model from a 2D reference. A vision-language model judge scored results across silhouette accuracy and structural completeness.

Hyra’s output scored higher after multiple rounds of revision. However, Hyra ran an iterative loop while Claude Code apparently used Goal mode. Whether the comparison controls for compute budget is unspecified. Therefore, this demonstrates recursive refinement benefits rather than definitive model superiority.

Future Goals for the Hyra Agent

Tencent positions Hyra as infrastructure rather than a consumer product. The goal is deploying it inside real R&D pipelines at scale. This converts environments into executable feedback loops that Hyra can optimize.

Future Hy models will co-evolve with this agent. Ultimately, the aim is automating the research loop itself. While benchmarks require scrutiny, the documented honesty about evaluator breakdowns makes this a rigorous release. The loop is simple, but that might be the point.