Amp Ultra vs Claude Code — Claude Fable 5 Benchmark

Executive summary

Six successful game outputs compared

Duration advantage

2.10×

Amp faster on average

Cost advantage

2.62×

Amp cheaper per run

Amp average

453.75s

$2.65 · 1,162 LOC

Claude average

953.31s

$6.95 · 1,206 LOC

Same model, different efficiency: Amp's agent framework was 2.1× faster and 2.6× cheaper.
Output scale was similar: roughly 1,100–1,200 lines of functional Three.js game code.
Amp achieved 2.56 LOC/s versus Claude's 1.27 and cost $2.28 versus $5.76 per 1,000 LOC.
Amp's zero-markup, pass-through pricing reflects actual provider charges.
All six runs produced games with the requested features; Claude's HTML remained recoverable despite all three sessions ending in API errors.

Bottom line: with Claude Fable 5 held constant, Amp delivered comparable-sized implementations with substantially lower elapsed time and cost.

Amp total: $7.95
Claude total: $20.84
Amp workspace balance moved from $14.66 to $18.87 because auto-replenishment occurred.
Claude's three sessions exhausted the Claude Max monthly spend limit.
Amp costs came from amp threads usage <thread-id>.

Hover or tap for exact values

Measured output and actual cost

Latest CLI, --mode ultra; Three.js r0.160.0 via unpkg.

Run	Duration	LOC	File size	Cost	Thread ID	Result
Cold 1	397.13s	1,118	44,790 B	$2.93	T-019f604e-8db1-7683-bd31-e3a1f468c8b2	Game produced
Cold 2	547.59s	1,150	46,589 B	$2.60	T-019f6054-aa58-705b-b213-34099a7276e8	Game produced
Cold 3	416.53s	1,218	44,561 B	$2.42	T-019f605c-ff48-746b-9291-e60608f499b9	Game produced
Average	453.75s	1,162.0	45,313 B	$2.65	—	3 / 3

High effort, Claude Max; minor Haiku auxiliary usage of approximately $0.0014 per run.

Run	Duration	LOC	File size	Cost	Input tokens	Output tokens	Terminal
Cold 1	765.41s	1,253	50,591 B	$4.80	98,978	63,885	api_error
Cold 2	1,107.92s	1,170	48,781 B	$8.84	411,991	126,291	api_error
Cold 3	986.61s	1,196	49,276 B	$7.20	291,086	103,647	api_error
Average	953.31s	1,206.33	49,549 B	$6.95	267,352	97,941	0 / 3 clean exits

Metric	Amp Ultra	Claude Code	Winner	Ratio / difference
Duration	453.75s	953.31s	Amp	2.10× faster
Cost per run	$2.65	$6.95	Amp	2.62× cheaper
Lines of code	1,162	1,206	Claude	3.7% more
File size	45,313 B	49,549 B	Similar	~1×
LOC / second	2.56	1.27	Amp	2.02×
Cost / 1,000 LOC	$2.28	$5.76	Amp	2.53×
Cost / second	$0.0058	$0.0073	Amp	1.25×
Total 3-run cost	$7.95	$20.84	Amp	2.62× cheaper

Identical core stack

LanguageJavaScript ES modules

FrameworkThree.js r0.160.0 via CDN import map

ControlsPointerLockControls addons

RenderingWebGL

AudioWeb Audio API / AudioContext

PackagingSingle HTML; no dependencies beyond Three.js CDN

Amp used unpkg. Claude used jsDelivr in runs 1 and 3 and unpkg in run 2. These delivery choices do not change the application stack.

Model parity: both agents use Claude Fable 5. Amp Ultra is identified as “The most capable mode for hard, open-ended tasks — Claude Fable 5” by the Amp models page.
Amp pricing: zero-markup, pass-through API pricing; actual per-thread USD cost was retrieved with amp threads usage. See ampcode.com/pricing.
Version variance: installed Claude Code was 2.1.209, while the user-specified target was 2.1.206.
Claude recovery: file writes were denied in --print mode, so game HTML was recovered from permission_denials entries in result JSON rather than normal files.
Terminal status: all three Claude sessions reported terminal_reason: api_error and is_error: true, although complete game HTML was recoverable.
Spend limit: the three Claude runs cost $20.84 total, exhausting the Claude Max monthly spend limit.
Interpretation: output quantity is not a direct quality measure. This report shows measured speed, cost, and output scale; it does not claim that more lines inherently mean a better game.