Wave Bench
run stats · from the Claude Code session transcripts

What each run cost.

Same prompt, same harness, one session per model. Bars share a scale within each panel; the longest bar is the largest value.

Active timeminutes
Fable 5.1 Claude Fable 5.1: 41 min 41 min
Opus 5.5 Claude Opus 5.5: 72 min 72 min
Sonnet 5 Claude Sonnet 5: 38 min 38 min
Sonnet 5.5 Claude Sonnet 5.5: 38 min 38 min
Output tokenstokens written
Fable 5.1 Claude Fable 5.1: 173K 173K
Opus 5.5 Claude Opus 5.5: 260K 260K
Sonnet 5 Claude Sonnet 5: 157K 157K
Sonnet 5.5 Claude Sonnet 5.5: 216K 216K
Estimated costUSD at list prices
Fable 5.1 Claude Fable 5.1: $14.41 $14.41
Opus 5.5 Claude Opus 5.5: $14.90 $14.90
Sonnet 5 Claude Sonnet 5: $10.27 $10.27
Sonnet 5.5 Claude Sonnet 5.5: $8.77 $8.77
Own codenon-blank lines
Fable 5.1 Claude Fable 5.1: 1,821 1,821
Opus 5.5 Claude Opus 5.5: 3,958 3,958
Sonnet 5 Claude Sonnet 5: 1,764 1,764
Sonnet 5.5 Claude Sonnet 5.5: 2,516 2,516
Cache readstokens re-read from cache
Fable 5.1 Claude Fable 5.1: 13.5M 13.5M
Opus 5.5 Claude Opus 5.5: 39.7M 39.7M
Sonnet 5 Claude Sonnet 5: 40.3M 40.3M
Sonnet 5.5 Claude Sonnet 5.5: 29.2M 29.2M
Simulation gridcells
Fable 5.1 Claude Fable 5.1: 25,600 25,600
Opus 5.5 Claude Opus 5.5: 48,000 48,000
Sonnet 5 Claude Sonnet 5: 12,000 12,000
Sonnet 5.5 Claude Sonnet 5.5: 22,400 22,400

All numbers

Fable 5.1Opus 5.5Sonnet 5Sonnet 5.5
Effortmediummediumhighhigh
Active time41.0 min72.2 min38.1 min37.9 min
Wall-clock time48.9 min77.2 min46.4 min45.8 min
Input tokens2,468270366234
Output tokens172,924259,556157,172216,112
Output tokens / active min4,2183,5954,1255,702
Cache reads13,452,18339,664,37940,309,71929,213,598
Cache writes190,273355,735256,127305,503
Estimated cost$14.41$14.90$10.27$8.77
Own code1,821 lines · 6 files3,958 lines · 16 files1,764 lines · 6 files2,516 lines · 14 files
Simulation grid160 × 160 · 2 m cells240 × 200 · 2 m cells120 × 100 · 1 m cells160 × 140
SolverStaggered-grid shallow water, upwind advection, wet/dry fronts, wave makerMAC-grid shallow water, semi-Lagrangian momentum, Flather offshore boundaryVirtual-pipes flux scheme with outflow clamping and sponge boundariesInertial shallow water on a staggered grid, implicit Manning friction, flux limiting

How these were measured

  • Active time drops gaps over 5 minutes; wall-clock time is first to last message.
  • Extracted from the Claude Code session transcripts of each one-shot run. Cost is an estimate at list API prices (cache writes billed at 1.25x input).
  • Cache reads dominate the token totals in agentic runs: every tool call re-reads the conversation so far.
  • Own code counts non-blank lines of the model's .js/.html/.css files, including its own tests and tools, excluding vendored three.js and generated bundles.
  • The Fable 5.1 session hit the output-token limit once and was told to resume; the Sonnet 5.5 session received some background-monitor notifications. Neither run got any additional guidance.
  • Effort differs: Fable 5.1 and Opus 5.5 ran at medium, both Sonnets at high.