Skip to content

Roadmap

What exists, what does not, and what is being built. This page is meant to be checkable — if something here is not true yet, it says so.

What exists today

  • run(jobs, deadline=...) across the OpenAI and Anthropic batch tiers, with a sync fallback that protects the deadline.
  • submit() / collect() and the Ticket — the same run split across processes: submit from a laptop or a CI step, save the ticket, collect from a cron the next morning. Nothing is orphaned at the provider when the caller dies.
  • quote(jobs, deadline=...) — pre-trade pricing with no API calls.
  • Receipts and settlements — list, paid, captured spread, and what a fallback left on the table, as arithmetic against published price sheets.
  • A Venue interface — the extension point. A venue is anywhere deferred work can run.
  • The Spread Board — GB power and carbon, marked daily.

What does not exist yet

Stated plainly, because a roadmap that reads like a feature list is a misleading one:

  • No queue-latency forecasting. Deadline risk is a fixed buffer, not a prediction. offpeak does not know how long a given venue's queue is; it watches the clock and falls back.
  • No cross-venue portfolio placement. Jobs route to the first venue that supports the model, not to the cheapest or fastest across a portfolio.
  • No carbon-aware scheduling. The Spread Board observes the grid; the scheduler does not read it.
  • A third venue settles: Google Gemini. 5 jobs, batch tier, 50.0% captured, zero fallbacks (receipt). Opt-in, like the rest.
  • Two more venues are written and cannot batch. Groq answers 403 not_available_for_plan; Mistral answers 402 ... enable billing via the console. Both drivers are exercised end to end and both have a receipt recording the attempt and what the sync fallback cost. Gemini was gated the same way until billing was enabled, so these are plan problems rather than code ones.
  • Two venues are written and have not run. DeepSeek is the first venue that is not a batch tier: it has no batch API and prices by the clock — half price outside 01:00–04:00 and 06:00–10:00 UTC on weekdays — so the driver holds jobs until the boundary instead of uploading them. Qwen on Alibaba Model Studio is an OpenAI-shaped batch tier with a region and a 24h–336h window. Neither has a receipt yet; both are opt-in, and the first sub-cent live run is what will verify them.
  • No venues beyond those. Spot capacity and off-peak windows on your own GPUs are interface-shaped but unwritten. A Groq batch driver exists in the tree, opt-in, excluded from the all extra and not in default_venues(). Its routing, request dialect, price rows and sync fallback are verified against the live API; its batch tier is not, and cannot be from here — Groq answers 403 not_available_for_plan to the entire Batch API on an unentitled key. The receipt for that attempt is 2026-08-23-groq-1.json: 24 real jobs, all of them through the sync fallback at list price, nothing captured.

The hosted desk

A hosted desk that does the forecasting, cross-venue portfolio scheduling, and SLA insurance at fleet scale — with payloads never leaving your perimeter — is being built by the same team.

The intended seam is the one already in the library: a desk would be selected per run, alongside the venues you already pass, so that moving from local scheduling to hosted scheduling is a keyword argument rather than a rewrite.

Not implemented

That parameter does not exist in the public API today, and nothing in this release accepts it. It is described here so the shape of the plan is legible — not as something you can call. The SDK and the deadline spec stay open, Apache-2.0, either way.

The spec

Deadline semantics are versioned separately in SPEC.md, so a second implementation can be written against them. Spec changes start as issues.