# Token and spend ranges for the challenge

These are planning estimates from selected entrant reports, not cohort averages or receipt-audited costs. Source: evaluation-matrix-complete.csv and spend-review.csv. Thirteen of 69 reviewed records contain quantified token claims; accounting differs. Estimates describe building and reviewing the submission, not ongoing production operation.

## Useful ranges

| Build pattern | Token scale in available samples | Financial scale |
|---|---|---|
| Smaller metered build | AbstractSync reports 15–20M total tokens. ProgramDesk reports 12.8M fresh-input plus output, excluding any unreported cached input. | AbstractSync reports ~$170; ProgramDesk $100 in credits. |
| Iterative, subscription-based build | Roughly 1–7B processed tokens is a useful planning band: SessionCanvas 963M, VibeBoard ~2B, CurtainCall 3.68B, Smolboard 3.95B, ProgramLoom ~4B, WeJustLearn ~6.5B. These are not identical accounting scopes. | Selected explicit examples report ~$100–$600 in credits, subscription allocation, or plan charges. This is not a measured price for the token band. |
| Intensive multi-agent build | Roughly 15–20B processed tokens: Programflow 14.814B, Chautauqua 18.07B, Eventloom 20B. | Programflow reports $400 in subscriptions with banked resets; Chautauqua reports three subscriptions plus $365.14 in evaluation API spend. No common cash-per-token conversion follows. |

For a similar ambitious multi-day attempt, budget roughly **$200–$1,000 for subscription access and coding credits**, and allow **another ~$200–$400 for paid evaluation if used**. This is an inferred planning envelope, not a promise that those plans provide enough allowance, and existing subscriptions may lower incremental cash cost. Selected heavy builds report **~$4,000–$12,000 of API-equivalent work**: Smolboard $3,709, Trackstage ~$4,000, WeJustLearn $4,875, ProgramCue $9,285, Chautauqua $11,612. These entrant-calculated equivalents are separate from cash paid and should not be added to subscription costs.

## What consumes the tokens

1. **Research and specification:** examining the brief, walkthrough, help docs, competitor flows, permissions and domain models. Trackstage used video-to-spec; OpenRostrum researched flows/buttons/features; SessionCanvas gathered assets and docs before phased implementation.
2. **Implementation and coordination:** scaffolding, database/auth, role-based portals, CFP/review workflows, scheduling, integrations; workers, orchestration and merges. Programflow recorded 112 root tasks and 46 spawned workers.
3. **Review, evaluation and repair:** adversarial diff review, tests, live evaluations, browser checks and repeated fixes. Programflow recorded 128 automatic reviews; WeJustLearn says review took the most effort; CallboardHQ reports at least 3.4M Claude tokens in its final evaluation alone. Separate evaluation API reports include $176.87 (OnFireside), $365.14 (Chautauqua), and about $400 (kms.r-s.workers.dev).
4. **UX refinement and delivery:** design iterations, manual product walkthroughs, mobile behavior, security checks, deployment, documentation and sometimes submission-video work. Programflow's scope includes launch-video concepts, so its total is broader than coding.

No defensible cohort-wide percentage allocation exists for these tasks. Two individual reports provide useful narrower breakdowns:

- AbstractSync's ~$170: ~$100 platform/scaffolding/UI/deploy previews, ~$20 research/planning subscription, ~$50 agentic coding and bug fixes. These are purchase buckets, not precise task-metered costs.
- Chautauqua's $11,612 API equivalent: ~$4,903 (42%) Opus planner/merge-trains/fleets; ~$2,994 (26%) Sonnet swarm workers/runner loops; ~$3,715 (32%) Fable interactive/meta-orchestration. These are model-and-role buckets with overlapping tasks.

## The cache explanation

Programflow: 14.814B processed = 14.504B cached input + 291M fresh input + 19.19M output (rounding differences). Chautauqua: 18.07B processed, 17.52B cache reads and 106.6M output. WeJustLearn: ~6.5B processed, ~6.4B cache reads and 9.9M generated tokens. Thus **about 97–98.5% of processed volume in these three detailed heavy examples is cached context reads**. This does not imply the same cost percentage. Smolboard separately reports rereading context at 72.7% of its API-equivalent bill and writing output at 6.5%.

Repeated context reading happens during research, implementation and review; it is not a separate product task. Billions of processed tokens do not mean billions of new code tokens.

## Suggested narration

“An ambitious attempt could process billions of tokens while costing its builder hundreds of dollars in subscriptions. In the largest reported runs, the count reached 15 to 20 billion—but roughly 97 to 99 percent was cached context being read again. The work included research, implementation, coordination, and repeated review and repair. Some builders estimated several thousand dollars of API-equivalent usage. That is a different number from what they actually paid.”