{
  "name": "Fabivo photo-to-furniture benchmark",
  "date": "2026-10-26",
  "codeRevision": "359e7f21369221d5c08e56ae76f54bd72158d6e6",
  "photos": 17,
  "prompt": "Model the furniture in this photo.",
  "reasoningEffort": "medium",
  "concurrency": 4,
  "sessionsPerPhoto": 1,
  "automaticRetries": 0,
  "selfCorrectionWithToolsAllowed": true,
  "averages": "Arithmetic means across all 17 photos, including every request and tool call within each session.",
  "inputTokens": "Uncached input plus cache reads plus cache writes, as reported by the provider through Pi.",
  "outputTokens": "Includes reasoning tokens.",
  "cost": "API list-price equivalent in USD, computed from per-request token logs, not the subscription bill. Terra reported no cache-write tokens; unreported writes could raise its estimate from $0.18 to about $0.21.",
  "transport": {
    "Claude Sonnet 5.5": "Pi Anthropic OAuth bridge",
    "Claude Haiku 5.5": "Pi Anthropic OAuth bridge",
    "GPT-6.1 Terra": "Pi Codex OAuth bridge"
  },
  "limits": [
    "Small internal set, one session per model and photo. No repeated trials or confidence intervals.",
    "Some examples from this set informed prior harness development. This is not a held-out evaluation.",
    "Two unusable reference cases were excluded before these runs. Every model saw the same 17 photos.",
    "Valid means passing the app checks, not proof of photo fidelity, structural strength or load capacity.",
    "Time measures the agent loop, including network and tool execution, not a person reviewing the plans or building the furniture.",
    "The first-slide overview contains five examples; the three full sheets contain all 17."
  ]
}