Changelog

Changes to the decision API that you can observe from the outside. Limits, models, request shapes and errors — newest first.

2026-11-09 — Five examples, and the page that runs them

/examples now carries five complete requests — support triage, comment moderation, lead qualification, expense categories and incident severity — each with the command to run it, the answer shape it must produce, and the keyless /api/decide variant where the questions allow one.

The bodies on that page are not transcribed from a terminal session. They are generated from the same definitions the test suite sends through the real handlers, with only the model provider stubbed; a body that stops being accepted fails the test instead of being published. What the page promises is reproducible is the request and the answer invariants — every declared label present, each distribution summing to 1, the model and request id echoed. The probabilities themselves are not, and the page says so.

  • Added /examples in the primary navigation, in the footer, in the sitemap and in llms.txt.
  • Each example names which endpoint it uses: POST /api/v1/decisions with a key, or POST /api/decide without one.
  • Where a keyless body cannot carry the whole request, the questions it omits are named rather than silently dropped.

2026-10-27 — Error codes, a published concurrency ladder, and /api/v1/usage

Refusals now carry a stable code beside the human sentence, so a client can branch on the reason instead of matching English. The code list and the per-plan limits are on /docs, and GET /api/v1/usage answers with the same figures for the key that asks.

Concurrency is now counted per API key and published: 8 requests in flight on Pro, 32 on Scale, 64 on Studio, 128 on Agency. It is counted in a Durable Object, one per key, because several isolates answer the same key at once and a per-isolate count is wrong by construction — measured on 2026-10-27, a caller sustained 8.7 requests a second, 286 calls in 33 seconds through the old per-address limiter.

  • Every error body is now { ok: false, code, error }. code is additive-only: a code is never reused with a different meaning. The full list is on /docs.
  • The three 429s are now distinguishable: quota_exceeded (today’s decisions), rate_limited (a burst), concurrency_exceeded (too many in flight). Before this release all three were the same 429.
  • 402 now distinguishes plan_ended (the term lapsed) from credit_exhausted (the credit the plan funds the model with is spent).
  • GET /api/v1/usage reports the plan’s limits, the requests in flight right now, and today’s high-water mark for that key.
  • Concurrency is counted and reported from this release. It is not refused until 2026-11-10T00:00:00Z: until then an over-limit call is answered, logged, and reported in concurrency.current. From that instant it is refused with concurrency_exceeded and retry-after: 1.
A stricter limit is announced at least one release ahead, which is why the counter ships before the refusal. The published figures are the ones the pricing page now shows beside each plan.

2026-10-27 — Clef-Omni, and a state limit that cannot truncate

A state that is too long for a model is now refused by this API before the call, with the model’s window and this state’s token count in the error. A 200 from /api/v1/decisions now means the model read the whole state, which it did not always mean before.

  • Added cloudflare/clef-omni (alias clef-omni) — Cloudflare, 65,529-token context, text and images, $0.15 per million input tokens. It reads the largest state this API accepts.
  • Corrected the published window for cloudflare/clef: it reads 16,384 tokens, not 65,537. Measured on 2026-10-27: 90,011 characters billed 16,246 tokens; 100,011, 120,011 and 200,011 characters all billed exactly 16,384, with the remainder discarded silently.
  • Every model is now capped locally in characters — Clef-Flash 120,000, Clef 80,000, Clef-Omni 200,000, Kev-4B 20,000 — and a state past its cap is a 413 that names the limit. Measured on the same day, cloudflare/clef-flash truncated a 150,011-character state and refused a 170,011-character one: the same request class, two different outcomes.
  • A 413 now carries the model, its token window and this request’s token count instead of the provider’s raw 422 body.
This release is about the answer being trustworthy, not about more capacity. If a state is refused, that is the API telling you the answer would otherwise have been computed from part of it.

2026-10-27 — Kev-4B, and a real model catalogue

The model list is now published in one place instead of being implied by the request parser, and it is enforced: a name that is not in it is refused rather than swapped.

  • Added jaredpalmer/kev-4b (alias kev-4b) — Jared Palmer's 4B open-weight Kev. 8,192-token context, text only, input billed at $0.042 per million tokens.
  • jev, jev-latest, kev and kev-latest remain aliases for this deployment's default model, and are now documented as aliases rather than left to be guessed. kev is not jaredpalmer/kev-4b; naming kev-4b is.
  • An image state sent to a text-only model now returns a 400 that names the model and the models that do read images, instead of reporting a zero image quota.
  • A state over 20,000 characters sent to jaredpalmer/kev-4b is refused locally before the call, because the provider answers an oversized Kev request with an error and no partial result.
Context windows are now on the record for every model. Clef-Flash reads 24,576 tokens; a longer state is truncated by the provider and billed at the full window. This was true before today — it is documented now because a 200 that silently lost the tail is worse than a refusal.

2026-10-26 — the body a Jev client already sends

The systemone endpoint stopped requiring a rewrite on the caller's side. Requests shaped the way jev and kev clients send them are accepted unchanged.

  • POST /api/v1/decisions serves the systemone contract: a state plus typed questions, answers keyed by question id.
  • questions accepted as an array of {id, …} or as an object keyed by id.
  • The jevapi single-label body accepted: instructions with labels, answered under the id category.
  • uol answers carry noul; score answers carry a float in [0, levels-1] plus per-level probabilities and a legend.
  • An edge block is reported by its cf-ray and cf-mitigated headers instead of a pasted challenge page.

2026-10-25 — images and payload size

Two limits that had been applied to every caller, paid keys included, were lifted for keyed requests.

  • Image states: base64 data URLs are accepted and passed through to the model. Remote https:// URLs are still refused, because the model does not fetch URLs.
  • Request body: the 8,192-byte default no longer applies to keyed requests. Paid keys accept bodies up to 8 MB and free requests up to 64 KB.
  • A 413 now reports both limits — what you sent and what is allowed — rather than a bare size error.

Frequently asked questions

How do I know a model change will not break my pipeline?

Name the model explicitly. cloudflare/clef-flash, cloudflare/clef, cloudflare/clef-omni and jaredpalmer/kev-4b are pinned ids; the aliases follow this deployment's default and can move. Every response reports the model that actually answered, so a diff of that field across runs is the check.

Do limits change without notice?

A limit that gets more permissive is announced here on the day it ships. A limit that gets stricter is announced at least one release ahead, except where the old limit was never real — in that case it is documented immediately and the entry says so.

Send a message