BDX Bench
AI coding models tested on one prompt. Real playable builds. Transparent scoring.
Round
Oct 2026
Models
9
Audited builds
11
Shared prompts
1
Audit dimensions
5
Showdown Score v2Re-audited at implementation level, not by keyword scan: 5 dimensions × 20 points, zero for absent features. Every verified defect stays published. How scores are produced
Featured builds
11 builds, one prompt. Every card runs the real artifact in place — play it, then read the score it earned.
Ranked by Showdown Score v2
6 enemy types with genuine role separation, not stat sticks
Five interlocking fire systems: heat, ignite, contagion, ground pools, death chains
PYRE — Burn the Horde
DeepSeek V4.1 Flash
Flamethrower cone: 175 px range, angled hit test, 42 DPS direct burn
INFERNO — Zombie Burner
Claude Haiku 5.5
4 enemy types with real behavioural divergence read inside the AI
Firebreak: Night Shift
GPT Luna 5.6
Genuinely working touch layer, verified via CDP multi-touch synthesis
INFERNO DEAD
DeepSeek V4.1 Flash
Proven continuous flame cone, measured 16.5 to 27.8 degrees with range
Cinderline
GPT 6 Sol
Real contagion cascades plus tank death explosions and 12% chain reactions
PYRO vs ZOMBIES
Muse Spark 1.3
Real continuous cone with a correctly wrapped angle helper
Emberfall
GPT Luna 6
A genuine radar minimap with real-time enemy and obstacle tracking
Space Bunny (superseded)
Space Bunny Free
A real 1/2/3 weapon system with three distinct mechanics
PYROCLASM: Zombie Inferno
Gemini 3.8 Flash
A well-tuned burn DoT with four secondary channels: slow, char, smoke, recolour
Zombie Fire Survival
Gemini Pro Agent
Audit credibility
Every Showdown Score v2 number comes from reading the build's source and playing it, not from scanning strings. Defects are published, not filtered out.
Implementation-level audit
Builds audited
11
Verified defects
48
Failing
22
Partial pass
18
Known issue
8
How we score
5 dimensions × 20 points = 100. A feature earns points only when it is implemented and reachable. On-screen strings, comments, and dead code score zero.
- Controls & Mobility14.4 / 20
- Combat Physics & Weapons12.0 / 20
- Content & Enemy Variety8.8 / 20
- Audio & Sound Design10.7 / 20
- Visual Polish & Game Feel11.7 / 20
Bars = cohort mean · ≥16 pass · 10–15 partial pass · <10 failure
Read the full methodologyHighest-severity findings
One verified failure per build, weakest-scoring build first. A benchmark that hides these is not worth reading.
Absolute zero audio
Failure6/6 probes empty and 0 AudioContext constructions at runtime. The game is completely silent.
Zombie Fire SurvivalGemini Pro Agent24/100
Advertised dash does not exist
FailureThe menu says Move & Dash; there is no dash, no i-frames and no shift handling.
PYROCLASM: Zombie InfernoGemini 3.8 Flash43/100
No flamethrower implementation
FailureThe word flamethrower appears exactly once in the whole file: the page title. There is no cone, no arc, no range code.
Space Bunny (superseded)Space Bunny Free49/100
One enemy archetype
FailureA single Zombie class with a size slider and no type field or AI branch.
EmberfallGPT Luna 651/100
| Build | Model | Score | Findings | Failing | Worst verified defect |
|---|---|---|---|---|---|
| EMBER DEAD | Space Bunny Free | 91 | 3 | 0 | No roguelite upgrade layerPartial pass |
| PYRE — Burn the Horde | DeepSeek V4.1 Flash | 80 | 4 | 2 | Zero touch supportFailure |
| INFERNO — Zombie Burner | Claude Haiku 5.5 | 63 | 7 | 0 | No touch inputPartial pass |
| Firebreak: Night Shift | GPT Luna 5.6 | 62 | 4 | 2 | Reachable pause hangFailure |
| INFERNO DEAD | DeepSeek V4.1 Flash | 61 | 4 | 2 | All four enemy types share one AIFailure |
| Cinderline | GPT 6 Sol | 58 | 4 | 2 | Knockback entirely absentFailure |
| PYRO vs ZOMBIES | Muse Spark 1.3 | 52 | 4 | 1 | touchcancel leaves the flamethrower latchedFailure |
| Emberfall | GPT Luna 6 | 51 | 4 | 1 | One enemy archetypeFailure |
| Space Bunny (superseded) | Space Bunny Free | 49 | 4 | 3 | No flamethrower implementationFailure |
| PYROCLASM: Zombie Inferno | Gemini 3.8 Flash | 43 | 5 | 5 | Advertised dash does not existFailure |
| Zombie Fire Survival | Gemini Pro Agent | 24 | 5 | 4 | Absolute zero audioFailure |
Severity labels follow the audit contract: failing = the build cannot do the thing it advertises · partial pass = implemented but wrong · known issue = cosmetic or unreachable. Each build's full finding list, including every line of evidence, sits on its card in Featured builds.
Global benchmark telemetry
Dataset scope9
Verified weights
1
Standard mini suites
9
Automated & manual
3
AI gateways & labs
Sep 2026
Methodology v1.0
Counts reflect the current dataset: 9 models · 1 benchmark · 3 providers.
Top models
Ranked by Showdown Score v2. Dimension columns are the best-scoring audited build for each model, scored out of 20.
| # | Model | Provider | Showdown | Controls | Combat | Content | Audio | Polish | Findings |
|---|---|---|---|---|---|---|---|---|---|
| 1 | Space Bunny Free2 builds audited | opencode | 91.0of 100 | 19 | 18 | 17 | 19 | 18 | 3no failures |
| 2 | DeepSeek V4.1 Flash2 builds audited | bdx-ai | 80.0of 100 | 16 | 17 | 14 | 17 | 16 | 42 failing |
| 3 | Claude Haiku 5.5 | claude-relay | 63.0of 100 | 15 | 16 | 7 | 12 | 13 | 7no failures |
| 4 | GPT Luna 5.6 | bdx-ai | 62.0of 100 | 15 | 13 | 11 | 12 | 11 | 42 failing |
| 5 | GPT 6 Sol | bdx-ai | 58.0of 100 | 15 | 11 | 9 | 11 | 12 | 42 failing |
| 6 | Muse Spark 1.3 | bdx-ai | 52.0of 100 | 12 | 12 | 9 | 8 | 11 | 41 failing |
| 7 | GPT Luna 6 | bdx-ai | 51.0of 100 | 15 | 9 | 6 | 10 | 11 | 41 failing |
| 8 | Gemini 3.8 Flash | bdx-ai | 43.0of 100 | 10 | 9 | 6 | 9 | 9 | 55 failing |
| 9 | Gemini Pro Agent | bdx-ai | 24.0of 100 | 11 | 6 | 2 | 0 | 5 | 54 failing |
Bar colour is the audit status for that dimension: lime = pass, amber = partial pass, red = failure. Findings are verified defects, published in full on each build. Other dimensions show as Not evaluated until measured.
Intelligence vs. price
Pricing and speed were not measured for these builds. See status
Not measured in this round
The eight verified builds were evaluated for game-build quality only. Price and speed charts will appear once vendor pricing and throughput measurements are published.
Category leaders
Zombie Flamethrower Showdown · game-build · the only category with measured data.
Leaders are the highest-scoring audited build per dimension, out of 20 points each. Other categories show as Not evaluated until measured.
Latest models
Capability trend
Single snapshot — trends need at least two rounds. Open trends
September 2026: 94 · 93 · 91 · 88 · 86 · 82 · 79 · 55
DeepSeek V4.1 Flash leads at 94.0, followed by Space Bunny Free at 93.0, GPT 6 Sol at 91.0, GPT Luna 5.6 at 88.0, Gemini 3.8 Flash at 86.0, Muse Spark 1.3 at 82.0, GPT Luna 6 at 79.0, and Gemini Pro Agent at 55.0. The next round will draw the first trend line.
Benchmark coverage
All benchmarkscoding · Showdown Score 0-100
8 evaluated builds · Showdown Score (manual game-build evaluation)
How we measure
Small, focused rounds with the same brief, clear checks, and judge review. Read the full rules before citing a number.
Read methodology →Fixed suite
Same brief and playability checks for every build in a round.
Hands-on scoring
Showdown Scores from direct review: 92 for Muse Spark 1.3, 88 for Gemini 3.8 Flash.
Open results
Per-build notes, provenance, and play links on every profile.