Is the $109 figure exact?

No — it's a model based on output artifacts and Opus 5.5 list pricing. Actual token counts depend on the specific prompts, tool use, and context window management. Treat as order-of-magnitude.

Does this include image/sound generation API costs?

No. The 783 images and 184 sounds were generated "through code" — likely via diffusion/TTS APIs called by the agent. Those are separate costs (typically $0.01-0.05/image, $0.01-0.10/second audio).

How does this compare to GPT-5 Soul for the same task?

At GPT-5 Soul pricing ($5/$15), same token volumes: ~$69 input + $69 output = ~$138 (assuming similar cache rate). Opus 5.5 wins on cache read ($0.20 vs ~$0.50) and output ($20 vs $15 is higher but Anthropic claims 40% less verbose output).

Can I run this today?

Opus 5.5 is available on Claude API. The 25-hour autonomous run requires a custom agent loop (not a single API call). See agent economics for architecture patterns.

Opus 5.5 long-running agent cost analysis

Updated September 27, 2026 · first published September 27, 2026

A demo revealed Opus 5.5 working autonomously for ~25 hours to build a complete Minecraft clone: ~100K lines of game code, 26K lines of tests, 1,000 automated tests, 775 files, 783 generated images, 184 generated sounds, and an in-game guide. This is the first public glimpse of a long-horizon coding agent at scale. Let's model the cost.

Estimating token volume

No official token counts published. We'll model from output artifacts:

Artifact	Est. tokens	Notes
100K lines game code	~3M output	~30 tokens/line avg
26K lines tests	~0.8M output	~30 tokens/line
775 files (structure, config)	~0.5M output	Boilerplate + imports
783 images (prompts + metadata)	~1.5M input	~2K tokens/image prompt
184 sounds (prompts + metadata)	~0.4M input	~2K tokens/sound prompt
In-game guide / docs	~0.3M output	Markdown + code snippets
Context / reasoning (25 hrs)	~10M input	Tool calls, observations, planning
Total estimate	~12.4M input / ~4.6M output

Cost at Opus 5.5 pricing

Token type	Volume	Rate	Cost
Input (uncached, est. 30%)	3.7M	$4.00/MTok	$14.80
Input (cached, est. 70%)	8.7M	$0.20/MTok	$1.74
Output	4.6M	$20.00/MTok	$92.00
Total			~$108.54

~$109 for a 25-hour autonomous agent run producing a shippable game. That's ~$4.36/hour of agent compute.

Sensitivity: cache hit rate matters most

Cache hit rate	Input cost	Total cost	Cost/hour
50%	$24.80	$131.60	$5.26
70% (baseline)	$16.54	$108.54	$4.34
90%	$8.88	$92.88	$3.72

At 90% cache hit (achievable with good context management), the run drops to $93 total / $3.72/hr.

Comparison: human developer

A senior dev building this solo: 3-6 weeks (~120-240 hours) at $100-200/hr = $12K-$48K. The agent run at $109 is 100-400x cheaper — but only if the output quality passes your release gate.

What this means for your agent budget

Long-horizon agents are affordable. Sub-$150 for a 25-hour complex build changes the ROI calculus for agent-first workflows.
Cache is the lever. Invest in prompt caching infrastructure — every 10% cache hit improvement saves ~$1.60 on a run like this.
Output tokens dominate. At $20/MTok, output is 85% of the bill. Models that are more concise (Anthropic claims 40% less verbose) directly cut cost.
Monitor per-run, not per-month. Agent workloads are bursty. Track cost per task completion, not aggregate monthly spend.