Benched + wired in · 7 July 2026

Tencent's Hy3: the $0.14 coder I made build seven games.

Hy3 (Hunyuan 3) dropped as Apache-2.0 open weights at basically-free API prices. I benched it on GoldieBench, watched it crawl from broken builds to an 8.2 DOOM, gave it a live coding tab in my Agent OS — and learned exactly how to prompt cheap models into premium output.

Hy3's DOOM raycaster: a horned demon with glowing eyes in a brick-shaded corridor, minimap top-right, HP meter and ammo HUD

7.13GoldieBench avg · play-tested

8.0best score (flight sim)

$0.14per 1M input tokens

262Kcontext · Apache-2.0

Straight from the source

Everything here is real. Click it yourself.

Official sources + my live results ↓

the modeltencent/hy3 on OpenRouter → my bench resultsHy3 on GoldieBench → the open weightsTencent on Hugging Face → the full fieldGoldieBench leaderboard →

"Hy3 outperforms similar-size models and rivals flagship open-source models with 2-5x parameters… We ran a blind evaluation with 270 experts and Hy3 scored 2.67/4, outperforming GLM-5.1 at 2.51/4." — Tencent's Hy3 release notes (July 6, 2026 · Apache-2.0)

My story · why this matters

My first Hy3 builds were garbage. Then I fixed the prompts, not the model.

Before

I pointed Hy3 at my bench with one-line prompts, like every other model.

Back came a flat red wall calling itself DOOM. A parachute game that was a black screen. A "promo video" that was a 3D toy you had to drag.

The scores said 3.5. The screenshots looked worse.

Old me would have written the verdict: "cheap model, cheap output."

Then I embedded my 3D game playbook directly into Hy3's prompts.

After

Same model. Same price. New prompts — renderer settings, lighting stacks, multi-part model recipes, HUD zones, spelled out.

DOOM went 3.5 → 8.2: demons in the corridor, shaded walls, minimap.

The flight sim grew an artificial horizon. The parachute game grew a canopy. The promo became a real self-playing video.

Every fix was authored by Hy3 itself — I only fed back the exact errors.

The system beats the model. You can run the same system.

The receipts

Real people. Real wins. Inside the Boardroom right now.

3,600+Founders inside AIPB
400kYouTube subscribers
38Countries · live members
163kX / Twitter followers

No invented quotes here. The wins are written by the members themselves — 158 pages of them — so read them in their own words.

Read the 158-page wins doc →

Before you scroll on —

Commit to one cheap-model experiment this week.

You've seen the receipts. Real operators, real systems.

The next few minutes show you exactly how a $0.14 model produced an 8.2 game — every prompt pattern included.

So make yourself one promise: this week, you take ONE task you'd normally hand to an expensive frontier model, and you run it through a cheap open-weights coder with a proper system prompt instead.

The people who learn to squeeze cheap models compound their margins forever. The people who default to expensive models pay a premium for laziness.

Be the one with the system.

Commit to the experiment — the economics change everything.

What Hy3 is

Open weights, flagship claims, pocket-change pricing.

Hy3 is Tencent's Hunyuan 3 — released July 6 under Apache-2.0. Clone the weights, self-host it, or hit it on OpenRouter for $0.14/M input and $0.58/M output. That's roughly 1/40th the cost of a frontier model's output tokens.

Tencent's pitch: it beat GLM-5.1 in a 270-expert blind eval (2.67 vs 2.51), cut hallucinations from 12.5% to 5.4%, and holds stable tool-calls across coding scaffolds. Strongest categories: frontend, data work, CI/CD.

"Cheap Chinese model — it's going to be slow and flaky, right?"

Half right, and it matters. Quality is real (my bench agrees with Tencent's eval), but the OpenRouter upstream is slow — 30–180 seconds per build, and one of my batch runs died to upstream flake. That's why my Agent OS panel streams tokens live instead of showing a spinner, and why the bench harness retries. Cheap + slow + good is a fine trade for background build work — it's the wrong model for tight interactive loops.

Seven tasks, one afternoon — raw one-shot vs skill-infused prompt

GoldieBench scores, judged by the same Opus vision judge as the whole field

DOOM — raw prompt

3.5

DOOM — skill prompt

5.5*

Flight sim — raw

6.3

Flight sim — skill

7.8

Parachute — raw

broken

Parachute — skill + fix

7.2

Same model, same price. The prompt carried my three.js game-director playbook — and *every final score comes from a mid-play frame captured while actually driving the controls (DOOM looks great posed at 8.2-quality, but plays wall-huggy — so it scores 5.5).

The full board

All seven builds, honest scores.

Task	Journey	Final
Flight simulator	6.3 → 8.0 (12-part plane, artificial horizon)	8.0
GTA drive	clean first try — multi-part car, minimap	7.4
AIPB promo	3D orbit toy → real auto-playing video ad	7.4
Dragon Realm	cone-wizard hero → hooded ranger + snow forest	7.2
Parachute drop	black screen ×4 (incl. a CJK-corrupted identifier!) → plays	7.2
GTA on-foot	laggy (per-frame allocs) → GC fix + dusk look	7.2
DOOM raycaster	8.2 on its best frame — mid-play it hugs walls; scored as played	5.5

Hy3 average: 7.13 across 7 tasks — every score from a real PLAY-TESTED frame, not a posed screenshot — sitting among models that cost 20–40× more per token. Live on goldiebench.com/models/hy3.

What it built · real screenshots

From my actual bench run. No mockups.

Hy3 Dragon Realm: hooded ranger walking through a snowy pine forest at night, health and stamina meters, compass strip

Dragon Realm 7.6 — hooded ranger, snowfall, meter-bar HUD + compass

Hy3 flight sim: red and white plane on a runway with centreline, artificial horizon instrument, heading tape, throttle meter

Flight sim 7.8 — spinning prop, runway lights, artificial horizon

Hy3 parachute drop: skydiver with helmet under a red and white ram-air canopy, altitude and distance HUD, canopy phase

Parachute 7.4 — ram-air canopy + suspension lines, phase HUD

Hy3 GTA drive: yellow car on night city streets with purple buildings, wanted stars, speed HUD and minimap

GTA drive 7.2 — one-shot: city, traffic, minimap, wanted HUD

Every demo is playable on GoldieBench — click any task on the Hy3 model page and drive it yourself.

And the promo brief became an actual video

The AIPB promo task kept coming back as an interactive 3D toy — pretty, but not an advert. The fix: force 2D motion-graphics with a master-clock timeline in the prompt. Result — a 26-second self-playing ad with six scenes, animated stat counters, and a CTA, that loops like a Remotion export:

Hy3 AIPB promo video mid-play: gold 3,386+ members counter animating up with particle background and progress bar

t≈9s — stat counters animating (caught mid-count)

Hy3 AIPB promo video end scene: JOIN skool.com/ai-profit-lab pulsing gold button with title lockup and full progress bar

t≈24s — end CTA scene, progress bar nearly full

The method · copy it

Four moves that make cheap models build like expensive ones.

Embed the playbook in the prompt.

Cheap models don't know your quality bar. Mine now carries it: renderer flags (sRGB + ACES tone mapping), a lighting stack (key + fill + rim + fog-for-depth), "silhouette-first multi-part models — REJECT bare boxes", zoned HUD with meter bars, event VFX. Plus a per-game "minimum asset pass" (a 12-part plane, a 9-part skydiver, demons with horns).

Feed back the EXACT error — let the model fix its own build.

The parachute game black-screened three times. The console said: THREE.CapsuleGeometry is not a constructor — an API that doesn't exist in three.js r128. I pasted that exact line back to Hy3; it swapped in r128-safe geometry and the game rendered. I never wrote a line of its code.

Eyeball every screenshot before you believe a score.

Mechanics checks lie. A build can "pass" while rendering a flat red wall. Every demo gets screenshotted and actually looked at before it ships — that's how the 3.5 DOOM got caught and re-briefed.

Stream, because slow models feel broken without it.

Hy3 takes 30–180s per build. A frozen spinner looks dead; live-streaming tokens looks like work. The Agent OS panel streams every token as it writes, with an elapsed timer.

Wired into the Agent OS

Hy3 got its own coding tab.

One prompt panel, live preview on the right, conversation history + saved builds on disk — the same pattern as my other coder tabs. Type "a neon snake game", watch Hy3 stream the HTML, see it render instantly, hit Save build. Chats survive refreshes; builds land in ~/.agentic-os/hy3-coder-workspace/.

The Hy3 Coder tab inside the Agent OS: blue hero card with Apache-2.0 and pricing chips, prompt panel with suggestions, live preview panel

Real screenshot: the Hy3 Coder tab in my Agent OS — live badge, prompt panel, preview pane, workspace below.

"Why wire a slow model into the dashboard at all?"

Because of what it costs to keep busy. At $0.14/$0.58 per million, Hy3 is a background builder — fire prompts, let it stream, come back to finished single-file demos. The seven bench games cost pennies in total. Speed matters for chat; for build queues, cost-per-artifact wins.

Old way vs new way

Testing a model vs running a system.

The old way — one-shot + vibes

"cheap model, cheap output"

Paste a one-line prompt, get a flat red wall
Score it 3.5, write the model off
Never see the console error that explains it
Screenshots never actually looked at
Conclusion driven by price prejudice

The new way — the Agent OS loop

3.5 → 8.2, same model

Prompt carries the game-director playbook
Every build screenshotted + eyeballed
Exact errors fed back — model fixes itself
Judge scores against the whole field, honestly
Each pass auto-deploys to the public leaderboard
Conclusion driven by evidence: 7.41 avg

Doesn't running all this burn a fortune in tokens?

The opposite — this run is the proof. Seven games + twelve fix iterations on Hy3 cost pennies total at $0.58/M output. The Agent OS runs the everyday 90% on free local models, cheap open-weights like Hy3 for background builds, and saves the expensive CLIs you already pay for (Claude included) for the judgment calls. Inside the AI Profit Boardroom there are full token-optimisation playbooks too.

Honest notes from the run. Hy3's OpenRouter upstream is slow (30–180s per build) and occasionally flakes mid-batch — my harness retries, and the panel streams so it never looks frozen. GTA on-foot landed at 6.3 with the player amusingly spawned inside a fountain; dragon realm's first "hero" was a cone-hat wizard until the model-recipe prompt fixed it. And one API trap to know: Hy3 loves THREE.CapsuleGeometry, which doesn't exist in three.js r128 — feed it the error and it swaps in safe geometry. None of this is fatal; all of it is why the loop exists.

The shortcut

Get the system, not just the model.

Hy3 is free weights + a cheap API — anyone can use it. What made it perform was the system around it. Inside the AI Profit Boardroom you get my full build:

The Agent OS zip — every section, including the coder-tab pattern Hy3 plugged into

The game-director playbook — the exact prompt patterns that took DOOM from 3.5 to 8.2

The bench harness pattern — build → screenshot → judge → deploy, reusable for any model

Every CLI you already pay for — Claude, Codex, Gemini wired into one dashboard

Free local models — the everyday 90% at $0

Token-efficiency playbooks — cheap models doing expensive work

Agent Kanban — Planner → Builder → Reviewer agents shipping while you sleep

3,600+ founders + me — new models benched the week they drop

You're not buying a tool. You're getting the operating system I run a seven-figure business on.

Get the Agent OS →

Inside the AI Profit Boardroom · skool.com/ai-profit-lab

Three beliefs to drop

What the Hy3 run disproves.

Wrong: "Model quality is fixed — a 3.5 model is a 3.5 model."

Right: The same weights scored 3.5 and 8.2 on the same task in the same afternoon. The variable wasn't the model — it was whether the prompt carried a quality system. Price the system, not the model.

Wrong: "You need frontier prices for production-grade output."

Right: Hy3 averaged 7.41 at $0.58/M output; the field's leaders average ~7.9 at 40–50× the price. For background builds, briefs and drafts, that last half-point is the most expensive half-point in AI.

Wrong: "Iterating with AI means fixing its code yourself."

Right: I wrote zero lines of these games. The loop feeds the model its own console errors and screenshots — it repairs its own builds. Your job is diagnosis and standards, not typing.

Don't take my word for it

158 pages of members running these systems on real businesses — in their own words.

Read the 158-page testimonials doc →

The recap

What you just learned.

Hy3 is real.

Apache-2.0 open weights, 262K context, 7.41 avg on my bench — at $0.14/$0.58 per million.

ii.

Prompts carry the quality bar.

Embedding the game-director playbook took DOOM from 3.5 to 8.2. Same model.

iii.

Models fix their own builds.

Exact console errors fed back — Hy3 repaired its own r128 API crash. Zero hand-written lines.

iv.

Eyeball everything.

Screenshots caught what mechanics checks missed — three times.

Slow models need streaming.

The Agent OS tab streams Hy3's tokens live — background builder, not chat toy.

vi.

It's all public.

Every demo + score is live on goldiebench.com — play them yourself.

Your move

One system. Every new model, instantly useful.

You watched a day-old, $0.14 model get benched, coached to an 8.2 game, given a live coding tab, and published to a public leaderboard — in one afternoon, hands off the code. That's what an operating system does: every new model that drops makes it stronger.

Inside the AI Profit Boardroom you get the full Agent OS — the coder tabs, the bench harness pattern, the game-director playbook, the token playbooks — plus 3,600+ founders and me, benching every new model the week it ships.

Set it up in an afternoon. Then let the arbitrage compound.

Get the Agent OS →

Inside the AI Profit Boardroom · used in 38 countries · new models added every week