Tencent Hy3 Tops OpenRouter: $0.14 Input Makes It the Cheapest Frontier MoE for Coding

Number One on OpenRouter at 10x Lower Cost

Tencent's Hy3 has reached the #1 position on OpenRouter's model rankings — and it did so at a price point that makes every other frontier model look expensive. At $0.14 per million input tokens and $0.58 per million output tokens, Hy3 costs a fraction of what competitors charge for comparable coding performance.

For comparison: GPT-5.6 Sol charges $5/$30, Claude Fable 5 charges $10/$50, and even the budget-tier GPT-5.6 Luna sits at $1/$6. Hy3's output pricing is nearly 10x cheaper than Luna and over 50x cheaper than Claude Fable 5 on output. A free tier is also available for low-volume usage.

Architecture: 295B Parameters, 21B Active

Hy3 uses a Mixture-of-Experts (MoE) architecture with 295 billion total parameters but only activates 21 billion per forward pass through top-8 routing across 192 expert modules. This is the key to its pricing: you get the knowledge capacity of a 295B model at the inference cost of a 21B model.

The 192-expert design with top-8 routing means each token activates only 8 of the 192 specialized sub-networks. This extreme sparsity allows Tencent to serve the model on significantly less GPU compute per request compared to dense models of equivalent quality. The context window extends to 262,144 tokens, making it viable for large codebase analysis tasks.

Coding Performance: What the Rankings Mean

Reaching #1 on OpenRouter is not a synthetic benchmark — it reflects real user preference across diverse tasks including coding, reasoning, and creative work. For coding specifically, Hy3 supports multiple inference modes:

• No-think mode — direct generation without chain-of-thought, fastest for simple completions
• Low CoT mode — brief reasoning steps before generating code, balanced speed and quality
• High CoT mode — extended reasoning for complex architectural decisions and debugging

The mode flexibility means developers can optimize cost per task type. Use no-think for boilerplate generation at minimal token cost, and high CoT only for complex logic where extended reasoning actually improves output quality. Tencent also emphasizes anti-hallucination mechanisms in Hy3, which reduces wasted tokens from incorrect code that needs regeneration.

The Cost Math: Hy3 vs the Field

Let us calculate what Hy3's pricing means for typical AI coding workloads. A standard coding session generating 50,000 output tokens (roughly 200 function implementations or refactors) costs:

• Tencent Hy3 — $0.029 (50K output tokens at $0.58/M)
• GPT-5.6 Luna — $0.30 (50K at $6/M)
• GPT-5.6 Sol — $1.50 (50K at $30/M)
• Claude Opus 4.8 — $1.25 (50K at $25/M)
• Claude Fable 5 — $2.50 (50K at $50/M)

At scale — say a team of five developers each generating 500K output tokens daily — the monthly difference between Hy3 and Claude Fable 5 is staggering: $435 vs $37,500. Even against the cheapest Western frontier option (GPT-5.6 Luna), Hy3 saves over 90%.

Tradeoffs and Considerations

Price is not everything. Before routing all coding workloads to Hy3, consider these factors:

• Data residency — Tencent's infrastructure is China-based; some organizations have compliance requirements that restrict data routing through Chinese servers
• Latency — geographic distance may add latency for developers outside Asia-Pacific regions
• English code quality — while strong on benchmarks, real-world English code generation quality should be validated on your specific stack
• Support and documentation — enterprise support infrastructure differs from Western providers

For teams where these tradeoffs are acceptable, Hy3 represents the most dramatic cost reduction available in frontier-quality AI coding. The MoE architecture proves that extreme parameter efficiency translates directly into pricing advantages that dense models cannot match at equivalent quality levels.

What This Means for the Market

Hy3's pricing puts pressure on every provider. When a model can top community rankings at $0.14/$0.58, the justification for $5-50 output pricing becomes harder to defend on quality alone. Expect Western providers to accelerate their own MoE deployments and continue the aggressive efficiency optimizations we are already seeing from OpenAI's Sol updates. The race to the bottom in AI coding costs is accelerating faster than most budget forecasts assumed.