Real head-to-head · same prompt, one shot
Claude Sonnet 5 vs Hy3
The agentic SWE frontier — 82% SWE-bench Verified, Dev Team mode. vs Tencent's open-weights coder — Apache-2.0, cheap, beats GLM-5.1 on frontend in Tencent's blind eval.
Head-to-head verdict: Hy3 wins 4–3.
What I tested — same prompt, two models
I run the same fixed prompt set through every new model the day it drops — same string, one shot, single HTML file out — and I score the result 0–10 on whether it ran, how close it hit the brief, and how good it looked. Below is what came out when I gave the exact same prompts to Claude Sonnet 5 and Hy3, side by side, on 7 shared tasks inside the Agent Operating System.
Both models were given identical prompts inside the Agent Operating System — no help, no iteration, no "best of N" tricks. I run each prompt once, save the HTML file the model produces, and score it 0–10 on whether it ran, how close it hit the brief, and how good it looked. The scoring is mine. The verdicts below are pulled from my source comparison guides at agentos.guide where I publish every score and the reasoning behind it.
Claude Sonnet 5 · Reach for it in Agent OS when the job is iterative, tool-using software engineering. For one-shot visual builds, GLM 5.2 (free) beat it 4-1 here.
Hy3 · Wired into the Agent OS as the 'Hy3 Coder' tab (chat + live preview + workspace) via OpenRouter. Bench built one-shot on the same prompts as the field; weak builds iterated by Hy3 itself (the model fixes its own builds).
Side-by-side on 47 shared tasks
Click any cell to play that model's actual one-shot attempt. Medals are derived from my 0–10 scores per task (highest = 🥇, second = 🥈, third = 🥉).
Task ↓
Claude Sonnet 5
Hy3
Game
Game
Game
Game
Game
Game
Page
Game
— not attempted —
Game
— not attempted —
Game
— not attempted —
Game
— not attempted —
Game
— not attempted —
Game
— not attempted —
Game
— not attempted —
Game
— not attempted —
Game
— not attempted —
Game
— not attempted —
Game
— not attempted —
Game
— not attempted —
Game
— not attempted —
Game
— not attempted —
Game
— not attempted —
Game
— not attempted —
Game
— not attempted —
Where Claude Sonnet 5 beat Hy3
The tasks where I gave Claude Sonnet 5 a higher 0–10 score on the same prompt — with the actual commentary from my source guides.
Doom
Game
Claude Sonnet 5 8.0
·
Hy3 5.5
(+2.5)
What I saw: Renders a clean raycaster maze with atmospheric red-lit walls, working minimap, HUD health/kills bar, crosshair and a monster visible at screen edge; solid feature set (hitscan shooting, chasing AI, damage flash, touch controls) makes it strong and shippable, though the wall shad…
Dragonrealm
Game
Claude Sonnet 5 8.0
·
Hy3 7.2
(+0.8)
What I saw: Strong, atmospheric night-time frozen realm — snowy terrain, pine forest, low-poly mountains and a moonlit sky read convincingly as Skyrim-esque, with a fully-modeled sword-wielding character and clean sword-draw/sheath system. Weak points: character proportions are a bit stiff a…
Aipbpromo
Page
Claude Sonnet 5 7.5
·
Hy3 7.4
(+0.1)
What I saw: 22KB · plays clean · plain
Where Hy3 beat Claude Sonnet 5
The tasks where I gave Hy3 a higher 0–10 score on the same prompt — with the actual commentary from my source guides.
Gtadrive
Game
Hy3 7.4
·
Claude Sonnet 5 4.5
(+2.9)
What I saw: Clean render with a readable yellow hero car (wheels, cabin, taillight), colorful blocky city, working HUD/minimap and speed at 101km/h shows live play. Solid and functional but visually flat-lit and generic — buildings read as bare boxes and lighting is dim, keeping it below the top tier.
Gtafoot
Game
Hy3 7.2
·
Claude Sonnet 5 4.5
(+2.7)
What I saw: Strong dusk-city atmosphere with a readable blocky hero (cap, jacket trim, gun), streetlights, crosswalks, and a pedestrian, plus clean HUD (ammo, health bar, wanted stars, crosshair, controls). Weak points: the world is fairly barren mid-frame, no visible buildings-as-cover in t…
Flightsim
Game
Hy3 8.0
·
Claude Sonnet 5 5.5
(+2.5)
What I saw: Clean readable HUD (SPD/ALT/VS, throttle bar, attitude ball, heading tape) plus a well-modeled shaded aircraft with red/blue wingtip lights, spinning prop, runway, tower, trees and lake — polished and clearly on-brief. Falls just short of the field's best: the attitude indicator …
Parachute
Game
Hy3 7.2
·
Claude Sonnet 5 6.0
(+1.2)
What I saw: Clean HUD (altitude, dist-to-target arrow, phase) and a nicely modeled skydiver with helmet visor, orange suit and boots reads well against drifting clouds. But mid-freefall the frame is mostly empty sky with no jungle/ground/clearing in view, so it feels sparse and generic rathe…
Strengths & weaknesses I logged
Claude Sonnet 5
Strengths
- 82.1% SWE-bench Verified — first model past 80% on real GitHub-issue repair
- Dev Team multi-agent mode + 1M context for repo-level agentic work
- Precision on hard logic — won the raycaster the open-weight field kept botching
Trade-offs
- One-shot creative-visual builds trail GLM 5.2 here (lost 4 of 5) — no iteration to catch its own bugs
- A temporal-dead-zone bug blanked its N-body orbit sim on the first shot
Hy3
Strengths
- Apache-2.0 open weights — self-host free, no lock-in
- Tencent's 270-expert blind eval: 2.67/4 vs GLM-5.1's 2.51, strongest on frontend / data / CI-CD
- Hallucination rate cut 12.5% → 5.4%; stable tool-calls across scaffoldings (<4% SWE-Bench variance)
Trade-offs
- Slow upstream on OpenRouter (30-90s per build) — fine for one-shots, sluggish for tight loops
- One-shot game builds can under-render (flat raycaster walls, unlit 3D) without an iterate pass
Pricing & context — the spec sheet
| Spec | Claude Sonnet 5 | Hy3 |
|---|---|---|
| Vendor | Anthropic | Tencent Hunyuan |
| Context window | 1,000,000 tokens | 262,144-token context window. Open weights (Apache-2.0) on HuggingFace / ModelScope / GitHub; benched here via OpenRouter. |
| Price | $3 / $15 per M ($2/$10 intro) | $0.14 / 1M input · $0.58 / 1M output |
| Pricing detail | $3.00 input / $15.00 output per million tokens; introductory $2.00/$10.00 through 2026-08-31. | Tencent Hunyuan 3 — open-weights under Apache-2.0, so free to self-host. On OpenRouter it is one of the cheapest capable coders: ~$0.14/M in, $0.58/M out (1 RMB / 4 RMB). Upstream can be slow (30-90s to first token), but per-token cost is negligible. |
| Release | 2026-06-30 | 2026-07-06 |
| Bench coverage | 47/47 scored · avg 7.01/10 | 7/7 scored · avg 7.13/10 |
The verdict — which should you pick?
Across 7 scored shared tasks, Hy3 averaged 7.13/10, beating Claude Sonnet 5's 6.29/10 by 0.84 points. Pick Hy3 when the build has to ship on the first prompt and you can afford the trade-offs in the comparison below.
If you only run one of these inside your stack, the head-to-head average above is the call. If you can run both, my honest play is to wire Claude Sonnet 5 and Hy3 both into the Agent Operating System and dispatch each from the kanban by task type — agentic software engineering — write / run / test / fix loops on real repos → Claude Sonnet 5, cost-sensitive coding + frontend design where open weights matter → Hy3. That's the same setup I run for the 3,600+ founders inside the AI Profit Boardroom.
FAQ — Claude Sonnet 5 vs Hy3
Which is better, Claude Sonnet 5 or Hy3?
On Goldie Bench, Claude Sonnet 5 averages 6.29/10 across the shared tasks, with 1 gold, 6 silver, 2 bronze overall. Hy3 averages 7.13/10, with 0 gold, 1 silver, 0 bronze. Hy3 wins the head-to-head 4–3.
How much does Claude Sonnet 5 cost vs Hy3?
Claude Sonnet 5: $3.00 input / $15.00 output per million tokens; introductory $2.00/$10.00 through 2026-08-31. Hy3: Tencent Hunyuan 3 — open-weights under Apache-2.0, so free to self-host. On OpenRouter it is one of the cheapest capable coders: ~$0.14/M in, $0.58/M out (1 RMB / 4 RMB). Upstream can be slow (30-90s to first token), but per-token cost is negligible.
What's the context window for Claude Sonnet 5 vs Hy3?
Claude Sonnet 5 has a 1,000,000 tokens context window. Hy3 has a 262,144-token context window. Open weights (Apache-2.0) on HuggingFace / ModelScope / GitHub; benched here via OpenRouter. context window.
When should I pick Claude Sonnet 5 over Hy3?
Pick Claude Sonnet 5 for: Agentic software engineering — write / run / test / fix loops on real repos; Repo-level reasoning across a 1M-token context (Dev Team multi-agent mode); Precise logic — raycasters, physics — where one-shot open models slip. The trade-off is the weaknesses we logged on the bench: One-shot creative-visual builds trail GLM 5.2 here (lost 4 of 5) — no iteration to catch its own bugs; A temporal-dead-zone bug blanked its N-body orbit sim on the first shot.
When should I pick Hy3 over Claude Sonnet 5?
Pick Hy3 for: Cost-sensitive coding + frontend design where open weights matter; Self-hosters who want an Apache-2.0 model they fully own; Anyone wiring a cheap capable coder into a live build panel (Agent OS Hy3 Coder tab). The trade-off is the weaknesses we logged on the bench: Slow upstream on OpenRouter (30-90s per build) — fine for one-shots, sluggish for tight loops; One-shot game builds can under-render (flat raycaster walls, unlit 3D) without an iterate pass.
How does Goldie Bench score Claude Sonnet 5 vs Hy3?
Every demo on this page was built by Julian Goldie inside the Agent Operating System — same fixed prompt for both models, one shot, single HTML file out. Each result gets a 0–10 score on whether it ran, how close it hit the brief, and how good it looked. The highest score on each task gets gold; second gets silver; third gets bronze. See methodology for full provenance.
Related comparisons
Other head-to-heads using the same scoring system:
Claude Sonnet 5 vs Fusion Hy3 vs Fusion Claude Sonnet 5 vs Hermes MoA Hy3 vs Hermes MoA Claude Sonnet 5 vs Claude Fable 5 Hy3 vs Claude Fable 5 Claude Sonnet 5 vs Grok Hy3 vs GrokFull model pages: Claude Sonnet 5 · Hy3 · back to the leaderboard
The same stack Julian uses
Run this stack yourself.
Every demo on this bench was built inside the Agent Operating System — one prompt, one shot, single HTML file out. The Agent OS, the prompts, the templates, the weekly walkthroughs and 3,600+ founders shipping with it every day all live inside the AI Profit Boardroom.
3,600+founders
258documented wins
38countries
$59/momonthly