claude-relay/claude-haiku-5-5 · rank 3 of 9
Claude Haiku 5.5
- Provider
- Claude Relay
- Family
- Claude
- Released
- Release date not published
- Closed weights
INFERNO — Zombie Burner
claude-haiku-5-5generated 2026-10-08No touch support
Flamethrower cone: 175 px range, angled hit test, 42 DPS direct burn
7 verified findings · no touch support · 5 audited dimensions
The live canvas starts when the window reaches your screen and stops when you scroll away.
Showdown Score
63.00/100
CapableShowdown Score v2strict source-code audit#3 of 9
Build report
The record this page's number is drawn from: one audited build, one score, and the provenance of both.
Implementation-level audit
- Generated
- 2026-10-08
- Audit method
- Showdown Score v2 · strict source-code audit
- Dimensions
- 5 × 20 pts
- Verified findings
- 7
- Build id
- claude-haiku-5-5
- Evaluator
- local-manual-eval
Suite: Zombie Flamethrower Showdown · round September 2026 · one manual run per model, so intervals are zero-width. Scoring rules, the withdrawn v1 method, and the limits of this round are in Methodology.
Evaluation breakdown
Where the 63.00 came from: 5 dimensions × 20 points. Each bar is the points awarded; under it is the verified evidence, and the defect that cost it what it lost.
1 pass · 3 partial · 1 failure
No capability recorded in the audit's implementation log for this dimension.
- Costs this dimensionNo touch input
Verified Flamethrower cone: 175 px range, angled hit test, 42 DPS direct burn
- Costs this dimensionNo weapon knockback
- Costs this dimensionOwn fire bombs damage the player
No capability recorded in the audit's implementation log for this dimension.
- Costs this dimensionStat-stick enemies
Verified Nine synthesized sound effects, all reachable, plus a looped fire-noise bed
- Costs this dimensionNo music
Verified Screen shake, persistent scorch and blood decals, night lighting, live HUD
- Costs this dimensionQuadratic zombie separation
- Costs this dimensionResize distorts saved decals
Bars: ≥16 pass · 10–15 partial pass · <10 failure · source audit 2026-10-08 · build claude-haiku-5-5
A feature earns points only when it is implemented and reachable — on-screen strings, comments, and dead code score zero. Full scoring rules in Methodology.
Prompt used
What was actually asked. The same text, verbatim and untuned, for every model in the round.
One prompt · 9 models · 11 audited builds
“make me a video game where i am killing zombies with fire and all please? a nice wonderfull game i can play for fun okay?”
| Benchmark | Score | Build | Updated |
|---|---|---|---|
| Zombie Flamethrower Showdown | 63.00 | claude-haiku-5-5 |
Showdown Score v2: strict source-code audit, five dimensions x 20 points. Community vote remains open in the interactive store. Every other benchmark dimension — reasoning, math, knowledge, pricing, throughput — is unmeasured for this build and is shown as such rather than estimated.
Failures found
Most benchmarks hide failures. These INFERNO — Zombie Burner defects were observed in the source and in play, worst first, and each one is why the score above is not higher.
0 failing3 partial pass4 known issue
No touch input
Partial pass
The viewport and fullscreen canvas imply mobile play, but the game cannot be controlled on a touch device.
Dimension costed · Controls
No weapon knockback
Partial pass
Flames and pools only reduce HP. Zombies are separated by overlap, not pushed by hits.
Dimension costed · Combat
Stat-stick enemies
Partial pass
Walker, runner and brute share one chase AI. They differ only in speed, HP, damage, size and attack cooldown.
Dimension costed · Content
No music
Known issue
Audio is one-shot effects, a looped fire-noise bed and groans, with no scheduled music track.
Dimension costed · Audio
Own fire bombs damage the player
Known issue
Lingering pools hurt the thrower at 9 HP per second with no warning ring, which punishes close throws.
Dimension costed · Combat
Quadratic zombie separation
Known issue
Every zombie pair is checked each frame with no spatial hash, so late waves of up to 80 zombies cost more per frame.
Dimension costed · Polish
Resize distorts saved decals
Known issue
Blood and scorch marks are stretched when the window is resized.
Dimension costed · Polish
Severity follows the audit contract: failure = the build cannot do the thing it advertises · partial pass = implemented but wrong · known issue = cosmetic or unreachable. A feature that only appears in a comment, an on-screen string, or unreachable code is not credited anywhere on this page.
Capability radar
Capability dimensions are Not evaluated for this build.
View radar values as a list
- ReasoningNot evaluated
- CodingNot evaluated
- MathNot evaluated
- KnowledgeNot evaluated
- VisionNot evaluated
- AgenticNot evaluated
Price and speed
0 of 5 measured
- Input / 1M tokens
- Not measured
- Output / 1M tokens
- Not measured
- Throughput
- Not measured
- Time to first token
- Not measured
- Context window
- Not measured
Pricing and speed were not measured for these builds. Values will appear here once vendor pricing and throughput runs are published — nothing on this page is filled in from a comparable model.
Score history
1 evaluated round
Rounds evaluated
1
Snapshot
63.00
Round date
Claude Haiku 5.5 has one evaluated round. A trend appears here only once a second independent round exists — a line drawn through a single measurement would imply a change that was never measured.
Related models
9 models · one prompt
Every model answered the same prompt. Scores are Showdown Score v2, so the numbers are comparable; the leaderboard adds the dimension-level evidence behind them.
Comments
Loading comments…
Community comments — stored on this server