# SYMBIONT: DYAD Issue 1 — honest cost and time ledger
**Compiled 2026-09-13 from files on disk and vendor transaction logs. Counts are generations that were paid for, including every experiment, dead end and re-roll. Shipped art is 24 pages plus a cover.**

Jeff's rule: "You just have to be honest. You gotta include all of the double ways as well." So this counts everything.

## Time
| | |
|---|---|
| First generation | 2026-09-03 (model head-to-head) |
| Print-ready and live | 2026-09-13 |
| Elapsed | **10 days** |
| Full or partial art passes | **6** (painted v1, v2, restyle v3, fix round 1, GPT Image 2.5 round, fix round 2) plus plates, cover and repairs |
| External blind reviews acted on | 4 |
| Claude Code sessions involved | 3 concurrent at peak |

The "90 seconds" that used to be on the page was the wall-clock time of ONE 24-page API batch. It was never the build time of the issue, and it is gone.

## Money
Unit prices used: Gemini 3.1 Flash Image 1K $0.067, 2K $0.101, 4K $0.151. Higgsfield credits $0.05 each, measured from the transaction log.

| Phase | Generations | Cost | Basis |
|---|---|---|---|
| 9/3 model head-to-head (Gemini, GPT Image 2, Qwen, Seedream, Nano Banana, Kling, Grok, Muse, Reve, Flux) | 29 | ~$3.15 | Higgsfield $0.65 measured; Gemini ~10 at $0.10; **fal share ~$1.50 is an ESTIMATE**, fal has no per-call export here |
| v1, painted style, 24 pages at 1K | 24 | $1.61 | files on disk |
| v2, 24 pages at 2K | 24 | $2.42 | files on disk |
| Plates, style tests, 4K resolution test | 7 | $0.76 | files on disk |
| Hero panels and two cover directions | 5 | $0.51 | files on disk |
| v3 restyle with plate chaining, 24 pages | 24 | $2.42 | files on disk |
| Fix rounds 1 and 2 (other session, Gemini), rejects included | 31 | $3.13 | files on disk; that session reported $2.93 |
| GPT Image 2.5 via Higgsfield (p14, p17, p19, 3 attempts on p19) | 7 | $0.85 | 17 credits, transaction log |
| Gemini Flash QA passes on the comic pages | ~30 calls | ~$0.20 | estimate |
| ElevenLabs voices for the read-along: first cast on multilingual_v2 (75 lines, 31 Rae retakes, splash, hymn auditions), then a full re-voice of all 82 lines on eleven_v3 after the Pro judge found mispronounced Spanish, a flat hymn and a boy voiced by a man; plus A/B and Terrell auditions (3 library voices added) | ~12,500 characters | $0.00 marginal | prepaid Pro plan, 369k characters unused this cycle |
| Read-along QA: casting, hymn and music judges; 6 rounds of Flash cold watch + sync pass on two cuts; Gemini 3.1 Pro cold watches and A/B judges | ~34 video/audio calls | ~$2.00 | estimate; Pro calls on 6.5-minute video cost several times Flash |
| Sound effects: 11 effects for 13 lettered SFX, 2 takes each, then 3 more takes for the 4 that failed (crowbar and slice sounds scored 2-4 until rewritten as foley) | 34 generations | $0.00 marginal | prepaid ElevenLabs credits; Gemini Pro judging included in QA row |
| Music | 1 instrumental (Voltage), after the first choice was rejected for lyrics | $0.00 | owned Suno catalogue |
| **Total** | **~171 images + a voiced read-along in two cuts** | **~$17.05** | of which ~$3.70 is estimated |

**171 images generated to ship 25.** About 7 generations per shipped page. The two full passes that were thrown away (v1 and v2, $4.03) and the rejected fix attempts are the "double ways" and they are in the total.

## What was NOT free and is not in the table
Human attention. Three sessions, four blind reviews, and Jeff's own reads across 10 days. No dollar figure is put on that here because it would be invented.

## Superseded claims, for the record
- "Art cost $2.42" (reader page tile): that was v3's generation batch alone. Removed.
- "Build time 90 seconds" (reader page tile): one API batch, not the issue. Removed.
- "First-pass rate 23 of 23": true of one batch, misleading for the issue, since most pages were regenerated more than once. Removed.
- "About ten dollars" (reader page close): my own running estimate before the fix rounds and the GPT round. Corrected to the ledger figure.
- "24 pages in ~90 seconds" (skill and plan): reworded to say what it is, one batch's API time.

## Read-along time, for the record
Four renders of about 10 to 15 minutes each were lost before one finished: a session restart killed one, `setsid` does not exist on macOS, an ffmpeg filter fed one output to two consumers, and a full-page splash overran the padded canvas. None cost money. All of it was wall-clock.

Six further full renders were stopped or superseded as QA found real faults: a stale cue sheet, a mask too faint, an inset left lit, a duplicated page view, a truncated ending, and the voice problems above. No money, all wall-clock.