Update: OpenAI's Decisions API hits public beta, pricing confirmed
The typed-decisions endpoint is now self-serve on GPT-6 Luna: $0.10/M input, no output fees, ~10x faster. Calibrate it yourself; Jev wins on long rubrics.
Copy markdown
What's new: self-serve public beta
OpenAI's decision endpoint has moved from limited release into open public beta. POST /v1/decisions takes your content plus up to 200 named questions and returns typed probabilities in one call — predicate (a 0–1 likelihood), choice (per-option odds plus a confidence field), or score (a weighted average over an ordered rubric). You threshold the numbers in code instead of parsing a chat reply, which turns classification, routing, moderation, and rubric grading into deterministic plumbing.
$0.10/M in, nothing out — and about 10x faster
It runs on GPT-6 Luna with a 1.1M-token context and bills input only: $0.10 per million tokens, with no output, cache-read, or cache-write charges. OpenAI clocks typed answers near 150ms versus roughly 1.6s on the Responses API. GA is expected 'in the coming weeks'; for now it's Luna-only and ZDR-eligible with prior approval.
The catch: calibrate it yourself
OpenAI ships no calibration data and tells you to set thresholds against your own labeled examples. Early public tests found predicate questions well-behaved, but one choice question's probabilities shifted when the option order changed — treat the output as raw signal to threshold, not ground-truth confidence.
A cheaper rival for repeated rubrics: Jev
TypeSafe's Jev decision model lists at $0.042/M input with free output — under half Decisions' per-token rate, and far cheaper when you reuse one long rubric across many items (an estimate pegs a cached-rubric workload near $97 vs $230 per million calls). Decisions wins on short, unique inputs and single-vendor ops and residency; Jev wins on volume.
Gemini 4 Argon reportedly nears wider access
Google is reportedly preparing a broader rollout of its Argon flagship — which launched gated to cyber-defenders first — while testing an internal variant nicknamed 'Carbon'. This is reported, not official, so hold the detail loosely; if you've been waitlisted for Argon, access may open up soon.