OpenAI Decisions API GA for Salesforce routing: the call

The call

OpenAI made the Decisions API generally available on 22 October, built on GPT-6 Luna through a dedicated endpoint, at $0.10 per million input tokens with no output-token charges (vendor, OpenAI). It reached general availability a little over three weeks after its 29 September preview, via a 6 October public beta (independent, Crypto Briefing). The watch we opened in our 1 October Field Note is closed: there is a price, a reference, and now general availability.

What it is, and what it is not

What it is. A typed decision endpoint. You send context and a bounded set of answers; it returns a predicate probability, a choice with probabilities and confidence, or a score (vendor). General availability adds up to 400 questions per request, twice the beta limit, image and JSON inputs, prompt caching, GPT-Live integration, Zero Data Retention for eligible customers, and US and EU inference residency (vendor). Crypto Briefing reports it as HIPAA eligible (independent).

What it is not. It is not a new model. OpenAI’s product lead Nikunj Handa says the team “reconfigured” Luna “to speed up the inference and lower the cost” (vendor, via TechCrunch via AI Weekly). It is not a price cut: the input price is the beta price. And it is not the accuracy leader. TechCrunch reports public benchmarks, unnamed in the coverage, ranking OpenAI’s decision model eighth and fourteenth, with Jev third and seventh (independent, via AI Weekly). OpenAI’s launch page gives no accuracy figure of its own.

The number nobody is quoting

Coverage leads with “up to 10x faster”. The line nobody quotes is the cache price. OpenAI’s GA page says cached input costs $0.01 per million tokens (vendor). At beta, cache reads and writes were not billed, which is what we recorded on 10 October. Crypto Briefing’s GA write-up still says they are free (independent). The vendor page wins until OpenAI says otherwise.

A case router is the cache-heavy shape: the same queue definitions and rules on every call, a short case body that changes. Above 64.4% cached, Decisions beats Jev’s uncached list price. Below it, Jev stays cheaper. Either way the bill is tens of dollars per million cases. A misrouted security case costs more than that in an hour. Accuracy on your own cases is the business case; the token price is the tiebreak.

Missing

What OpenAI has not published on its GA page, as of 23 October:

• An accuracy figure on any named benchmark.
• Rate limits and an SLA for the Decisions endpoint.
• Cache-write prices for 30-minute and 12-hour writes: the page links to a pricing guide we could not confirm.
• A written answer on the cache price change from beta, given that GA coverage still reports caching as free.
• Who counts as “eligible” for Zero Data Retention.
• Regional and residency premiums at GA.
• Fallback behaviour when a question is refused or confidence is low, and how that is billed.
• A median latency: the only p95 is a customer’s.

Who owns what

Latency is their job. Routing accuracy is ours.

In Salesforce, the call sits behind case creation. A record-triggered Flow, or an Apex callout through a Named Credential holding the API key, sends the case subject and body with your queue list as the choices. Write the choice and the confidence back to case fields, let Omni-Channel route on the field, and send anything under your confidence threshold to a human triage queue. For volume, publish a Platform Event and call out asynchronously, so a slow response never blocks case save. Keep the queue definitions in a stable prefix so the cache does its job.

Commercial decision: one vendor for every decision, or the cheapest accurate router per job. On today’s evidence, decide per job and keep the callout swappable.

Monday morning

1. Step 1. Pull 500 closed cases with their final queue. Strip personal data. If you cannot, start with our 240-case synthetic set at /data/salesforce-case-routing-testset.csv. Owner: service operations lead.
2. Step 2. Run Jev and Decisions in shadow. Our harness at /data/decision_routing_harness.py already has a decisions caller. Score accuracy, p95 latency, and cost per case. Owner: Salesforce architect.
3. Step 3. Measure your cached share. Count prefix tokens against total, then price it at $60.00 − $54.00s per million cases. Owner: Salesforce architect.
4. Step 4. Put the Missing list to OpenAI in writing. Rate limits, SLA, cache pricing, and ZDR eligibility. Date the reply. Owner: IT procurement.
5. Step 5. Set the confidence threshold for human triage. Pick it from the shadow run, not from a launch post. Owner: service lead.

We move this to WATCH if published rate limits or an SLA cannot carry your peak case volume.