Hy3 vs Claude Sonnet 5
The Apache Agent vs The Proprietary Coder
July 7, 2026 · 10 min read
Two Very Different Philosophies
Tencent Hy3 (July 6, 2026) is an open-weight 295B MoE under Apache 2.0 — 21B active parameters, 192 experts, $0.20/$0.80 per 1M tokens. Built in under 90 days and hardened through 50+ internal Tencent products, it prioritizes reliability: hallucination rates dropped 57%, commonsense errors halved, and tool calls work consistently across scaffolds.
Claude Sonnet 5 (June 30, 2026) is Anthropic's proprietary Sonnet-tier flagship — "the most agentic Sonnet yet." It uses adaptive thinking, has a 1M context window (4× Hy3's 256K), and launched at $2/$10 per 1M tokens (through Aug 31), then $3/$15. Anthropic doesn't disclose parameter counts. It's available on Anthropic API, Bedrock, and Vertex AI.
Coding: Sonnet 5 Leads Decisively
| Benchmark | Hy3 | Sonnet 5 | Δ | Winner |
|---|---|---|---|---|
| SWE-bench Verified | 78.0 | 85.2 | +7.2 | Sonnet |
| SWE-bench Multilingual | 75.8 | 78.3 | +2.5 | Sonnet |
| SWE-bench Pro ★ | 57.9 | 63.2 | +5.3 | Sonnet |
| Terminal-Bench 2.1 | 71.7 | 80.4 | +8.7 | Sonnet |
| FrontierCode v1 | — | 38.8 | — | Sonnet* |
The SWE-bench Verified gap (+7.2) and Terminal-Bench lead (+8.7) are the most telling for developers. Sonnet 5 is meaningfully better at real GitHub issue resolution and CLI operations. But note: Sonnet 5's 63.2% Pro costs $10/1M. Hy3's 57.9% costs $0.80/1M. That's 91.7% of the Pro capability at 8% of the price.
Agents & Search: Near-Ties Where It Matters
| Benchmark | Hy3 | Sonnet 5 | Δ | Winner |
|---|---|---|---|---|
| BrowseComp ★ | 84.2 | 84.7 | +0.5 | ~Tie |
| MCP Atlas (Public) | 79.1 | — | — | Hy3* |
| DeepSearchQA | 91.0 | — | — | Hy3* |
| OSWorld-Verified | — | 81.2 | — | Sonnet* |
* Not published by the other vendor
Hy3's MCP Atlas score of 79.1% — #3 overall, #1 open-weight — is a genuinely impressive result. The BrowseComp near-tie (84.2 vs 84.7) is the most actionable signal: on Anthropic's own agentic search benchmark, a $0.80/1M open-weight model is 0.5 points behind a $10/1M proprietary model.
Reasoning & STEM
| Benchmark | Hy3 | Sonnet 5 | Δ | Winner |
|---|---|---|---|---|
| GPQA Diamond | 90.4 | 91.1 | +0.7 | ~Tie |
| HLE (with tools) | 53.2 | 57.4 | +4.2 | Sonnet |
| HLE (no tools) | 37.0 | 43.2 | +6.2 | Sonnet |
| USAMO 2026 | 72.0 | — | — | Hy3* |
The Economics: 12.5× Price Gap
| Metric | Hy3 | Sonnet 5 (intro) | Sonnet 5 (std) |
|---|---|---|---|
| Input /1M tok | $0.20 | $2.00 | $3.00 |
| Output /1M tok | $0.80 | $10.00 | $15.00 |
| Output Multiplier | 1× | 12.5× | 18.75× |
| Context | 256K | 1M | 1M |
| License | Apache 2.0 | Proprietary | Proprietary |
For 100M output tokens/month: Hy3 costs $80. Sonnet 5 costs $1,000 (intro) or $1,500 (standard). And there's the tokenizer caveat: Sonnet 5's updated tokenizer produces 1.0-1.35× more tokens for the same input, meaning your effective cost per character increases after the introductory period ends.
The Self-Hosting Factor
Claude Sonnet 5 cannot be self-hosted. Hy3, under Apache 2.0, can run on your own hardware — 295B total / 21B active fits a 2× DGX Spark configuration. For regulated industries, this alone may be the deciding factor.
10-Point Verdict
Which One?
Choose Sonnet 5 if:
- Coding is paramount. +5.3 Pro, +7.2 Verified, +8.7 TB 2.1 are decisive.
- You need 1M context. For long-horizon coding sessions across many files.
- You want managed infra. Bedrock, Vertex, Anthropic API — zero DevOps.
- You're in the Anthropic ecosystem. Claude Code, prompt caching, computer use.
Choose Hy3 if:
- Cost matters at scale. 12.5× cheaper at intro, 18.75× after August.
- You need to self-host. Apache 2.0, fits consumer hardware.
- You're building agents. MCP Atlas 79.1%, BrowseComp near-tie, DeepSearchQA 91.0%.
- You want predictable token economics. No tokenizer surprises, no pricing cliffs.
The Bottom Line
Claude Sonnet 5 is the better coder. Hy3 is the better value agent. At 12.5× the price, Sonnet 5 delivers +5.3 Pro points — a premium justified for code-heavy workloads. But for agent workloads, the BrowseComp near-tie suggests the open-weight vs proprietary gap has nearly closed. Pick your weapon.
Test Both on Real Code
20+ LLMs on CodingFleet. Run Hy3 and Sonnet 5 side-by-side.
🚀 Try on CodingFleet →Sources: Tencent · Anthropic · System Card · Vellum · Emergent.sh. Vendor-reported scores unless noted.