1,194,890,784 input and output tokens across all included sessions.
Engineering economics
Agentic development, with the meter visible.
Cinch publishes the consumption behind the proof of concept: infrastructure footprint, token volume, model mix, delivery evidence, and the caveats needed to interpret each number honestly.
Tokens processed1.19B97.03% of input served from cache
Pilot run rate$45–60estimated / month
Delivery window91commits
Model roles2build + review
Actual invoices—not connected
Development consumption
Context is expensive. Reuse changes the curve.
1.19B tokens moved through the Codex context window during the first recorded app-engineering window. The dominant economic fact is not the headline total: 1.16B input tokens were served from cache.
97.03% of input reused previously processed context.
The smaller portion of input that was not served from cache.
Includes 708.11K reasoning-output tokens.
Cached input · 97.03% Uncached input · 2.97%
Model mix
Different roles, separately visible.
GPT-5.6 Sol
981.89M82.17% of processed tokens
gpt-5.6-solCodex automated review
213M17.83% of processed tokens
codex-auto-reviewDaily volume
A short, intensive build window.
Session start dates group the counters. They do not represent active hours or continuous human work.
- Aug 8484.96M
- Aug 9612.88M
- Aug 1097.05M
Delivery evidence
Consumption needs an outcome denominator.
Token volume becomes useful only when it can be interpreted beside shipped changes, tests, documentation, operating evidence, and quality gates. These Git numbers describe the same recorded window.
Small, named checkpoints across product and operating work.
32 feature · 16 test · 24 documentation
Across 360 unique paths; volume, not a quality score.
feature32
test16
documentation24
fix8
build3
other8
No labor conversion: Cinch does not translate tokens, commits, or changed lines into “developer hours saved.” That would imply a precision the evidence cannot support.
Infrastructure cost
A small pilot footprint, built for real operating behavior.
The documented August 2026 budget envelope is $45–60 per month before meaningful egress and optional monitoring growth. It covers the working topology—not a hypothetical single-container demo.
Always-on web availability, sync compute, managed Postgres, persistent replica storage, and wildcard TLS.
$0$60 alert
Stateless web pair kept running for blue/green releases and host-failure tolerance.
Persistent pilot sync node; horizontal sync availability remains a planned milestone.
Managed Postgres with an always-on replication floor, logical replication, and recovery controls.
Encrypted persistent sync replica with automatic snapshots and a tested rebuild path.
Cookie-free Vercel deployment with its own source, release identity, and rollback history.
Monitoring treats the top of the pilot budget envelope as an early-warning signal, not a hard cap.
Development cost basis
Consumption is measured. Dollar attribution is deliberately withheld.
These sessions ran through ChatGPT-managed Codex. Applying API token prices would create an API-equivalent scenario, not an actual invoice. The portfolio therefore publishes token consumption and model mix now, and reserves a dollar total for a source that can support it.
No per-session dollar charge is present in the local event logs.
Measured consumption is shown without pretending it is provider spend.
A dated rate-card snapshot could model equivalent API cost as a clearly separate estimate.
Official references: Codex JSONL usage events · workspace analytics API · Codex pricing and plan behavior
Methodology & caveats
Every number carries a provenance label.
The snapshot can be regenerated without exposing the underlying conversations. Its data contract contains aggregate counters, dates, model identifiers, and Git statistics only.
Codex token counters
Cumulative per-session counters are differenced and aggregated by model. Raw prompts, responses, local paths, and identifiers are never emitted.
Git delivery history
Commit metadata in the same time window supplies delivery context. Commit and line counts describe change volume—not quality or human-equivalent effort.
Infrastructure envelope
The application runbook records the deployed pilot topology and an August 2026 public-rate budget range. It is not a provider invoice.
Actual billing feed
Fly, Neon, monitoring, and supporting-provider invoices are not yet normalized into a project-scoped cost export.
Included
Aggregate token fields · model labels · session counts · UTC dates · Git summary
Excluded
Prompts · responses · session IDs · local paths · user data · secrets · private infrastructure identifiers
See the operating model