OpenAI splits classification into a Decisions API priced at 10 cents per million input tokens

In Simon Willison's first test, he sent an image file named two-pelicans.jpg with the question "Does this image contain any mammals?" and got back {"type": "predicate", "name": "evaluation", "probability": 0.0} ^[12]. Pelicans are birds, so the model was right ^[12]. The whole answer is a few tokens, and OpenAI charges nothing for output ^[6]. Prompt length and attachment size set the bill.

The yes/no type returns a probability ^[1]. The caller picks the cutoff. Moving it changes how often a router escalates without changing the price of a call ^[6].

OpenAI says the endpoint runs about ten times faster than its Responses API ^[24]. That figure describes OpenAI's own comparison payload. It transfers to your queue if your current classifier runs through Responses and spends most of its time generating output, such as a JSON wrapper or a reasoning preamble ahead of the label. A classifier already constrained to one short answer has less generation time to cut.

Jev got there first. The Decoder calls the launch likely OpenAI's response to the "decision models" trend Jev started in mid-September ^[8]. Willison's 6 October post reported the release, following an announcement at DevDay the week before ^[9]. He wrote that the API shape "is very similar to Jev, at least conceptually" ^[4]. Both offer the same three question types ^[2].

Jev charges 4.2 cents per million input tokens. OpenAI charges 10 ^[11]. OpenAI is about 2.4 times the price ^[19]. A billion input tokens a month costs $100 on OpenAI and $42 on Jev ^[20].

The premium buys two things. "Unlike Jev, the new gpt-6-luna decision model supports image input in addition to text," Willison wrote ^[7]. OpenAI says the API evaluates text, images or both, and lists damage detection in photos among its use cases alongside inquiry routing and document classification ^[3][10]. It also supports zero-data retention and HIPAA-compliant use in the US and Europe ^[13]. A text-only routing queue with no health data uses neither feature, and Jev is the cheaper place to run it ^[19]. The OpenAI endpoint is also in public beta with one model ^[5].

Moving between the two looks cheap. Willison built his llm-openai-decisions plugin by having GPT-6 Astra read OpenAI's documentation, modeled on the plugin he already had for Jev ^[14]. I'd run the same labelled sample through both and compare error rates and latency on my own data before committing either way.

The tier change is a budgeting problem. OpenAI cut its paid tiers from five to three: Build, Launch and Grow ^[15]. Organisations move up automatically when total credit purchases reach the next threshold ^[16]. The monthly limits are $500, $5,000 and $200,000 ^[17]. At Decisions prices, Build covers 5 billion input tokens a month and Launch covers 50 billion ^[21]. The step from Launch to Grow is 40x ^[22]. A workload that needs anything in between has to buy credits up to the Grow threshold first ^[16]. The Decoder's report does not list the thresholds or say whether reaching a limit stops requests. On 3 October, Willison published a post titled "We're going to need default hard budget caps on pretty much everything" ^[18].