offpeak

Deadline-priced inference. Same model, same tokens, a later deadline — for half price.

OpenAI, Anthropic, and Google all sell batch inference at 50% off list price. Almost nobody uses it, because no API lets work say it can wait: every token runs "now" by default, and the batch workflow — build a file, upload, poll, download, match results back up — is enough friction that urgency gets bought by accident.

offpeak gives your code one new argument.

import offpeak

jobs = [offpeak.job("claude-haiku-4-5", f"Summarize:\n\n{doc}") for doc in docs]

results = offpeak.run(jobs, deadline="06:00")   # done by 6am, at batch prices

print(offpeak.receipt(results))

OFFPEAK SETTLEMENT ────────────────────────────
jobs      1,000 (1,000 ok, 2 sync fallback)
sla       1,000/1,000 met
venues    anthropic:batch 1,000
tokens    12,410,332 in · 3,104,551 out
list      $27.93
paid      $14.02
captured  $13.91 (49.8%)
prices    snapshot 2026-08-21 — override via offpeak.prices
───────────────────────────────────────────────

Documentation · Quickstart · Spec · Roadmap

What it does

• Know the price before you spend it. quote(jobs, deadline=...) prices a run against the published sheets with no API calls and no key — list versus batch, per venue, plus what the wait is worth.
• One argument, not a workflow. run(jobs, deadline=...) handles batching, submission, polling, collection, and result matching across providers.
• Deadlines are guarded, not hoped for. If a batch hasn't landed by the time the remaining window shrinks to a risk buffer, offpeak cancels and re-runs the stragglers synchronously at list price. You state the deadline; it gets met.
• Every run settles a receipt. List cost, paid cost, captured spread — arithmetic against public price sheets, not estimates.
• Your keys, your perimeter. offpeak talks directly to the providers with your own API keys. There is no proxy and no third party in the data path.
• Zero-dependency core. Provider SDKs load only via extras.

Install

pip install "offpeak[all]"        # OpenAI + Anthropic venues
pip install "offpeak[anthropic]"  # or one provider
pip install "offpeak[openai]"

Venues use the standard environment variables (OPENAI_API_KEY, ANTHROPIC_API_KEY), or pass a configured client: OpenAIBatch(client=my_client).

Five more venues are in the tree and opt-in — each wants its own key and its own extra, and is passed to run() explicitly rather than routed to by default:

from offpeak.venues import DeepSeekClock, QwenBatch

results = offpeak.run(jobs, deadline="06:00", venues=[DeepSeekClock(), QwenBatch(region="intl")])

DeepSeek is the first venue that is not a batch tier: peak is 01:00–04:00 and 06:00–10:00 UTC on weekdays, everything else is off-peak at half the peak rate, and the driver holds a job until the clock is cheap rather than uploading it anywhere. Neither DeepSeek nor Qwen has a live receipt yet.

The free quote

What is the wait worth? Ask before you spend anything — quote() makes no API calls and needs no key.

python -m offpeak quote --model gpt-5.6-luna --input-tokens 800 --output-tokens 200 --jobs 5000

OFFPEAK QUOTE ─────────────────────────────────
jobs      5000 across 1 venue(s)
deadline  2026-08-21 21:11 PDT (24.0h out)
tokens    4,000,000 in · 1,000,000 out

  openai:batch      5000 job(s)  list $2.00  batch $1.00  save $1.00 (50.0%)

list      $2.00   (run now, synchronously)
batch     $1.00   (run by the deadline)
save      $1.00 (50.0%)
risk      deadline is inside the 24h batch window — the SLA rests on the sync fallback, which pays list
basis     input explicit; output explicit
prices    snapshot 2026-08-21 — estimate only, not a bill
───────────────────────────────────────────────

From Python, offpeak.quote(jobs, deadline="06:00") takes the same jobs you would pass to run(). Token counts come from the job where it knows them (metadata={"input_tokens": ..., "output_tokens": ...}, or max_tokens as an output ceiling) and are a labeled chars/4 estimate where it does not — every figure reports its provenance in basis, and a quote with no output signal is marked a floor, not an estimate.

If you do know roughly what the model will write, say so and get a priced number instead — quote(jobs, deadline=..., assumed_output_ratio=0.25), or metadata={"expected_output_tokens": 300} on a single job. Those quotes are marked EST, distinct from a floor. Both are opt-in: absent one, offpeak assumes nothing on your behalf.

Deadlines

Deadlines are how software says "this can wait" — the full semantics live in SPEC.md.

How a run works

1. Jobs are grouped by venue (claude-* → Anthropic Message Batches, gpt-*/o* → OpenAI Batch) and submitted at the batch tier — 50% off list.
2. offpeak polls the venues, backing off while the window is long.
3. When remaining time reaches the risk buffer (default: 15% of the window, clamped to 1–10 minutes), unfinished jobs are cancelled and re-run synchronously so the deadline holds. Set fallback="none" to report them instead.
4. Results come back in input order, each with a per-job Receipt; offpeak.receipt(results) settles the run.

results = offpeak.run(
    jobs,
    deadline="06:00",
    fallback="sync",       # meet the deadline at list price if the batch is at risk
    risk_buffer=600,       # seconds held in reserve (optional)
)

Receipts and prices

Receipts are computed against a bundled snapshot of public list prices (batch = 50% off list, as published). Providers change prices — verify and override at runtime:

import offpeak

offpeak.prices.register_price("my-fine-tune", input_per_m=4.0, output_per_m=16.0)

Unknown models settle with cost = None rather than a guess. The sheet also carries what the venues charge for urgency — get_fast_price(), urgency_spread() — and flags list prices that are promotional, with the date and the price they decay to: promo_decay("gpt-5.6-sol") is (1.25, 1.5) after 2026-11-21.

What this is (and the roadmap)

offpeak is the open client and spec for a simple claim: intelligence has a time value. A large share of AI work — embeddings, evals, backfills, report generation, daily agent runs — has no human waiting on it, and the venues already sell that patience at 50% off. This library is the missing workflow.

The token side is wider than the headline discount. Patience sells at 50% off — a 2.0x spread — and haste is priced too: hold the model and venue constant and gpt-5.6-sol costs $8.00 / $40.00 per 1M on OpenAI's fast tier against $2.00 / $10.00 on its batch tier, a 4x intra-venue urgency spread for the hour alone (source; sol's standard rate is promotional at least through 2026-11-21, and both tiers are defined off it, so the ratio outlives the prices). It is data, not prose: offpeak.prices.urgency_spread("gpt-5.6-sol") returns 4.0.

The Spread Board marks the same claim against open grid data daily — GB power and carbon plus CAISO SP15 and ERCOT Houston peak/off-peak spreads, alongside the published token spreads. ERCOT Houston marked 3.94x on the session of 2026-08-20; a venue charges 4x for the same impatience.

The roadmap follows the same interface upward: more venues (Google batch, spot capacity, off-peak windows on your own GPUs), queue-latency forecasting instead of a fixed risk buffer, portfolio placement across venues, energy- and carbon-aware scheduling with per-job receipts. The venue interface (offpeak.Venue) is deliberately the extension point — a venue is anywhere deferred work can run.

A hosted desk that does the forecasting, cross-venue portfolio scheduling, and SLA insurance at fleet scale — payloads never leaving your perimeter — is being built by the same team. The SDK and the deadline spec stay open, Apache-2.0.

Contributing

Issues and PRs welcome — see CONTRIBUTING.md. Spec changes start as issues against SPEC.md.

License

Apache-2.0 © Offpeak