The Spread Board

offpeak rests on a claim: intelligence has a time value. The token side of that claim is already settled, and it is wider than the headline discount.

• Patience is priced at 50% off. OpenAI and Anthropic both publish batch tiers at half price: a flat 2.0x spread for work that can wait.
• Haste is priced too. Hold the model and the venue constant and read the same sheet across its urgency tiers: gpt-5.6-sol is $8.00 / $40.00 per 1M tokens on OpenAI's fast tier against $2.00 / $10.00 on its batch tier — a 4x intra-venue urgency spread, on both legs, for the hour alone. Source: developers.openai.com/api/docs/pricing.

The promo caveat

gpt-5.6-sol's standard rate is promotional — the sheet says it runs at least through 2026-11-21, after which list is $5/$30. Fast and batch are both defined off that list (2x and 0.5x), so the promo moves the dollars and leaves the ratio: the 4x is the durable figure, the prices are the perishable ones. Both are in the SDK rather than in prose — offpeak.prices.urgency_spread("gpt-5.6-sol") returns 4.0, and promo_decay() returns the step-up the date will bring.

The Spread Board marks the same claim against the other side of the trade: the grid the compute runs on. Each session it records what power and carbon actually did between the evening peak and the trough that follows, from open, keyless sources. The grid's spread is not published anywhere — it has to be observed — and in the session of 2026-08-20 the ERCOT Houston hub marked 3.94x between its evening peak and the trough that followed, against the 4x a venue charges for the same hour of impatience.

The two sides of the grid do not move together, which is the point of marking both. On 2026-08-18 and 08-19, CAISO's carbon ran cleaner at the evening peak than in the trough — spreads of 0.76x and 0.73x — because the sun that serves the California evening has set by midnight. Cheap hours are not automatically clean hours, and a board that only recorded price would have implied otherwise.

→ Read the board

How it works

Two passes over the same session, run by a scheduled workflow:

Pass	When	What it records
quote	19:00Z	The session ahead — carbon forecast, day-ahead power
mark	06:30Z	The session just finished — carbon actuals

A session runs 16:00Z–07:00Z: it opens with the 17:00 BST evening peak and closes after the 00–05 BST trough, so one span carries both windows the board compares.

Output lands on the board-data branch — nightly/BOARD.md plus the raw JSON per session — because main is protected and would reject a scheduled bot push.

Settled runs are a different ledger

BOARD.md observes; it spends nothing at any venue. Runs that actually executed and actually billed go in nightly/SETTLED.md on the same branch, written by tools/settle_report.py from the receipts in receipts/ — never by hand — and published by a manual workflow, because a settlement is a deliberate act.

Every settled row carries its scale, and the column is not decoration. A few dozen jobs proving the mechanics end to end and a production book are both real settlements and are not the same evidence. A ledger that lets a reader confuse them is doing marketing rather than accounting, so the scale is printed before the money is.

Queue latency is a third table

offpeak abandons a slow batch on a fixed risk buffer — 15% of the window, clamped to 1–10 minutes — because nobody had the number. nightly/QUEUE.md collects the number: a couple of tiny jobs submitted daily at each venue's declared window, timed from submit to completion. The desk publishes a daily summary of its queue observations here — the observations themselves are private.

It spends real money, so it is a separate table from the board for the same reason SETTLED.md is, and it is capped at one cent per run, enforced inside the desk's tooling rather than in a runbook.

The summary is percentiles over completed sessions, not single rows: p50/p90/ p99 and max per venue, over the trailing 7 days, the trailing 30 days, and all time. A row marked censored was still open when a run stopped waiting and was cancelled — a lower bound, never a completion time — and counts toward the attempt total but not toward any percentile.

First observations, 2026-08-23: Anthropic Message Batches landed a 2-job book in 85s and OpenAI Batch in 2m26s — 0.098% and 0.169% of the 24h window each declares. Groq is absent, and not by choice: its Batch API answers 403 not_available_for_plan without an entitled plan, so there is nothing to time and its seven completion-window strings remain unconfirmed.

Sources

• Carbon — NESO carbon intensity, GB, keyless.
• Power, GB — Octopus Agile day-ahead unit rates, GB region C, keyless.
• Power, US — CAISO SP15 and ERCOT Houston day-ahead hourly, via gridstatus, keyless.
• Carbon, US — EIA-930 hourly generation by fuel for the CAISO and ERCOT balancing authorities, through the EIA Hourly Grid Monitor. Needs a free API key; without one the column records itself unavailable and nothing else changes.
• Tokens — the published price sheets, not a measurement: OpenAI and Anthropic.

All are public and free. The board costs nothing to run and spends nothing at any venue.

Honest limits

• Four zones, unevenly covered. GB carbon and GB power are half-hourly and complete. CAISO SP15 and ERCOT Houston are day-ahead hourly prices, not settled real-time ones.
• US carbon is derived, GB carbon is measured. NESO publishes an intensity; EIA does not. The US columns are computed from EIA-930's hourly generation mix times EIA's own CO2 coefficients and fleet heat rates — every input published, the product an estimate, and marked "basis": "derived" in the record so it is never confused with a measurement. It counts generation, not consumption: imports and the carbon already stored in a battery are outside what the method can see, and the share of generation EIA files under "other" is reported per session rather than averaged in.
• US carbon arrives about two days late, and every mark goes back for it. A session's carbon spread compares its evening peak against the trough of the following morning, 00:00–05:00 local — the next calendar day. So marking the session of day D needs EIA published through D+1, and EIA itself runs about a day behind. At 06:30Z the column is therefore never fillable for the session just ended: not sometimes, always. Each mark pass re-fetches the last few sessions whose carbon is still missing and rewrites those records, so the column fills itself roughly two days later and the projection republishes the whole history. A column that has not arrived yet is recorded as unavailable with the reason, never as zero.
• The token column is published, not observed. The 2.0x and the 4x are read off price sheets; only the grid columns are measurements. A published number and a marked one are different kinds of claim, and the board should not blur them.
• Observation, not advice. The board records what the grid did. It does not forecast, and offpeak does not currently schedule against it — the SDK routes on published token prices alone.
• Carbon actuals lag roughly two hours. The 06:30Z mark clears the 00–05 BST trough comfortably; the tail of the session can still be sparse.
• A dead source costs its column, not the run. Legs degrade independently, and a session where both sources are down is recorded as unavailable rather than guessed at.