OpenAI Decisions API for first-pass review: does it earn a seat in your doc workflow?

What it is

OpenAI Decisions is a stripped-down API endpoint for one job: forcing a model to return a bounded answer — a yes/no, a pick from a list, or a score — instead of free text. The Decisions API evaluates text, images, or both and returns typed answers about 10x faster than the Responses API. Get the probability that a condition is true, a choice from a fixed set, or a score against a rubric. It runs on a single underlying model and sits alongside the normal chat API, not in place of it.

The next-work-session test

You have 40,000 documents from a production. On Monday, your team needs a first pass that splits "likely privileged / likely responsive / junk" before an associate ever opens one. Today that first cut either eats contract-review tool budget or burns Responses API output tokens on prose explanations you throw away. Decisions lets you send the doc plus a fixed rubric and get back one tag plus a confidence score, which your review platform ingests as a column. The win is narrow but real: cheaper, faster triage before human eyes.

Pricing

OpenAI has not published a dedicated public rate card page I could find, but third-party coverage of the beta announcement is consistent. OpenAI's Decisions API bills classification and routing at $0.10 per million input tokens with no output charge, making the actual cost roughly one-fifth of what standard Luna pricing implies. OpenAI opened a dedicated Decisions API in public beta on October 6, running only on gpt-6-luna.

Two caveats before you build a budget on that. First, the no-output-fee saving isn't a universal discount — one analysis notes that for short items with no shared prompt it is about 9% cheaper, because you stop paying for output tokens, and that long shared rubrics can cost 4x more than Responses because there's no cache discount. Second, there's no official SLA tied to the beta pricing — treat the rate card as "today's beta price."

What we'd actually use it for

Three things, honestly:

That's it. We would not use it to decide anything that ends up in a privilege log without a human sign-off.

Limits

It is a classifier bolted onto a general model, not a legal-grade review tool. A few things to be clear-eyed about:

• It doesn't explain itself. By design, you get a label and a probability, not reasoning. If your QC process depends on "why did the model flag this," you'll need a second Responses API call on the flagged subset — which erases some of the cost advantage.
• Confidence ≠ correctness. A neatly typed answer can still be wrong, biased, or based on incomplete state. What is publicly known about OpenAI Decisions API. The safest current summary is that the OpenAI-branded Decisions API is not a normal, generally available endpoint — it's beta, and the usual beta risks (rate limits, breaking changes, no enterprise contract muscle) apply.
• Privilege is contextual. Thread position, custodian role, and matter context all matter. A per-document classifier won't catch a privileged chain where only the parent email names counsel.
• No built-in audit trail for discovery defensibility. You'll need to log prompts, rubrics, versions, and outputs yourself if you want to defend the methodology later.
• Model lock-in to one tier. You can't swap in a reasoning model for the hard calls through this endpoint.

For the broader "where it fits vs. legal AI tools" framing, the Artificial Lawyer write-up is worth a read — it makes the point that this is a complement to, not a replacement for, the Harveys and LegalOns of the world.

Try it if

• You run (or your e-discovery vendor runs) a document review pipeline and you control the orchestration layer.
• You have a technical team that can wire an API into Relativity, Reveal, or a custom review platform.
• You want to benchmark cost-per-doc for first-pass triage against your current provider.
• You're happy running a beta endpoint on non-production data first.

Skip it if

• You're a small firm without engineering support — wait for a legal-tech vendor to wrap this for you.
• You need defensible, explainable output for every decision in the pipeline.
• Your matter is small enough that off-the-shelf review tools already handle it inside existing retainers.
• Your clients or regulators require data residency guarantees the beta doesn't yet contract to.

---

This content is for informational purposes only and is not legal advice. Confirm confidentiality, privilege, and jurisdictional rules before using any AI tool with client matters.