---
type: analysis
slug: decision-apis-are-repeating-the-chat-completions-playbook-and-once-again-the
title: "Decision APIs are repeating the chat-completions playbook, and once again the request shape will port while the probabilities will not"
description: "Perplexity's local Decider and OpenAI's general-availability Decisions API follow the path chat completions took: a common request shape, then a local server that speaks it. The last round showed that matching JSON does not mean matching behaviour. With decision models the gap is calibration."
published_at: 2026-10-27T14:00:00.000Z
topics: ["openai"]
source_urls: ["https://docs.perplexity.ai/docs/decisions/decider-for-mac","https://community.perplexity.ai/t/you-can-now-run-pplx-decider-flash-v1-2-locally-on-your-mac-without-an-api-key/6413","https://community.openai.com/t/decisions-api-is-now-generally-available-with-new-updates/1408934","https://www.infoq.com/news/2026/10/openai-devday-2026","https://vercel.com/changelog/perplexity-decider-v1-2-models-now-available-on-ai-gateway"]
author: "Andrey Reshetnikov"
author_profile: https://newruntime.com/owner-profile.md
---
# Decision APIs are repeating the chat-completions playbook, and once again the request shape will port while the probabilities will not

Perplexity's local Decider and OpenAI's general-availability Decisions API follow the path chat completions took: a common request shape, then a local server that speaks it. The last round showed that matching JSON does not mean matching behaviour. With decision models the gap is calibration.

## Summary

Perplexity released Decider for Mac, a menu-bar app that runs its 4B pplx-decider-flash-v1.2 model on Apple Silicon. It serves the same request shape as Perplexity's hosted Decisions API at 127.0.0.1:8080, with no API key and no usage bill. A day earlier, OpenAI moved its own Decisions API out of beta. It added caching, requests of up to 400 questions, and gpt-6-luna pricing with no charge for output tokens. Together the two releases make "ask a typed question, get back a probability" its own API primitive, with both local and hosted versions.

## What happened

Decider for Mac downloads a 5.8 GB q8_0 GGUF build of pplx-decider-flash-v1.2 on first launch. It needs macOS 13 or later, 16 GB of RAM and about 6 GB of free disk. The app accepts text and images, sent as state plus named questions, and returns probabilities and confidence for yes/no, choice and score questions. Perplexity's examples are labelling private files, a pre-commit hook that scans for secrets, and sorting screenshots. Perplexity says moving to the hosted API means changing the URL and adding an authorization header. The hosted service can run the larger pplx-decider-v1.2 model. Perplexity does not publish accuracy, latency, licence terms, or any claim that thresholds tuned locally will match hosted results.

On October 22, OpenAI's developer community announced general availability of its Decisions API. The release adds native GPT-Live integration for voice, prompt caching, native JSON inputs, per-question image instructions and requests of up to 400 questions, which OpenAI says is double the previous limit. For gpt-6-luna, input costs $0.10 per million non-cached tokens and $0.01 per million cached tokens, and output is free. InfoQ's DevDay 2026 recap lists the Decisions API alongside GPT-6.1 Sol, computer use in the Agents API and cloud environments for Codex. That placement makes it part of OpenAI's agent stack, not a one-off product. None of these sources includes independent latency or quality measurements.

## What the archive adds

Earlier archive coverage fills in the background. According to Vercel's changelog, Perplexity Decider V1.2 27B and Decider Flash V1.2 4B were already on Vercel AI Gateway as multimodal "decision models", filed under a separate decision modality. That means a gateway vendor had treated decisions as their own model type before the Mac app shipped. OpenAI's Decisions guide (d1) says predicates return the probability that a condition is true. Choice and score answers return a distribution plus a separate confidence field, and the guide tells developers to use labelled examples from their own application. The beta announcement (d2) said the API returns typed answers about 10x faster, a claim truncated in our copy, and that it charges nothing for output tokens. Perplexity's changelog (d4) describes Decisions as producing probabilities that software can use directly, without parsing generated prose, and supports routing or escalating to its agents. The precedent is in d5: llama.cpp ships a llama-server that starts "a local OpenAI-compatible server". That is how chat completions became the default local interface.

## Trend signal

The pattern is older than this month's releases. One vendor defines an endpoint shape, it becomes the shape everyone codes against, and local runtimes copy it so that developers can switch between a laptop and the cloud by changing a base URL. Chat completions went this way. The /v1/chat/completions route became the interface that llama.cpp, gateways and dozens of hosts implemented. Decisions are now going through the same process faster. OpenAI and Perplexity use the same product name and the same basic idea (typed questions in, probabilities out). Vercel has already made it a gateway modality, and Perplexity now ships a localhost server for it. Our reading is that "decision" is becoming the third standard primitive after generation and embeddings. It is cheap and fast, it leaves no prose to parse, and it fits agent routers, guardrails and pre-commit hooks.

## Why now

1) Agent stacks need cheap, bounded judgements: routing, escalation and guardrails. Perplexity's changelog (d4) explicitly ties Decisions to routing and escalation to its agents, and InfoQ places OpenAI's Decisions API among its Agents API and Codex updates. 2) Economics: OpenAI charges nothing for output tokens and, according to its beta post (d2), the API answers typed questions about 10x faster. Parsing generated prose is slow and costly at high request volumes. 3) Privacy demand: Perplexity pitches the Mac app at content that "can't leave your Mac", such as customer exports and logs containing tokens. 4) Precedent shows that local compatibility works as a distribution channel. llama.cpp's OpenAI-compatible server (d5) proved that a local endpoint in a hosted shape pulls developers toward the hosted product.

## Connections to other events

The earlier round predicts where this one breaks. Local OpenAI-compatible servers did make portability work at the level of the request. Code ran unchanged against llama.cpp or a hosted model. The behaviour behind the shape still differed. Tool-call formats, stop handling, and whether log-probs were supported at all varied between servers, and many teams found out only after switching. Decision APIs carry a sharper version of that problem, because the whole output is a number someone has put a threshold on. A pre-commit hook that blocks at p > 0.8 on a quantised 4B model has nothing to do with p > 0.8 on the hosted 27B model. Perplexity's own wording claims compatibility of request shape and nothing about compatible probabilities. OpenAI's guide tells developers to calibrate on labelled examples. Perplexity's "just change the URL" therefore holds for the code and fails for the decision logic. A team that prototypes thresholds on the Mac app and then moves to the hosted service has effectively changed classifiers without noticing.

The history also predicts a second step. Once a shape is the default, local runtimes and gateways implement it, so a generic /v1/decisions route in third-party servers is plausible. The OpenAI and Perplexity schemas are not shown to be identical, though, so gateways will probably normalise between them, just as they did for chat-completions variants.

## Why it matters

For builders, the local app is a good way to try decision workflows on sensitive data: secret scanning, labelling customer exports, triaging logs. Testing costs no money and the data stays on the machine, if Perplexity's on-device claim is accurate. Treat any threshold you tune as tied to one model, quantisation and version. Store it next to the model identifier and re-run your labelled set whenever you change URLs. Buyers comparing OpenAI's gpt-6-luna pricing with Perplexity's hosted tier should not assume one vendor's 0.7 means the same as the other's. Compare precision and recall on your own data at the operating point you will actually use. Write a thin adapter between your application and any decision endpoint, as teams eventually did for chat completions, so that changing provider means changing the adapter and a calibration file, not your business logic.

## What happens next

- **At least one widely used third-party local runtime or gateway (e.g. llama.cpp server, Ollama, or another gateway besides Vercel) ships a route explicitly compatible with a vendor Decisions API.** (6 months, 35%). Watch for: Release notes or a merged PR adding a /v1/decisions-style endpoint returning per-question probabilities.
- **Vercel AI Gateway, or another multi-provider gateway, lists OpenAI's Decisions models under the same decision modality as Perplexity's Decider, which would show a normalised cross-vendor schema.** (6 months, 50%). Watch for: A gateway changelog entry adding gpt-6-luna (or a successor) as a decision model next to pplx-decider-v1.2.
- **Perplexity publishes guidance or tooling on calibration differences between local pplx-decider-flash and hosted pplx-decider, such as threshold mapping or a parity evaluation.** (12 months, 30%). Watch for: A docs page or changelog that addresses differences in local versus hosted probabilities, not only URL and auth changes.

## Limits and uncertainty

Nothing in the evidence measures accuracy, latency or calibration for any of these models, so the size of the local-versus-hosted gap is our inference, not a measurement. It might be small. Perplexity may have trained the flash and full models to produce aligned probabilities, and it has not said either way. We do not know whether the OpenAI and Perplexity request schemas match closely enough for a single client to call both. The comparison to chat completions may also overstate how fast this spreads. Decision endpoints are a narrower market, and the flash model's licence, which the sources leave out, could stop other runtimes from bundling it. The Mac app is limited to Apple Silicon machines with 16 GB of RAM, which narrows its reach.

## Skill to build now

Probability calibration for classifier endpoints. Exercise for this week: take 150 labelled items from your own data, for example commits with and without real secrets. Score them with Decider for Mac and, if you have access, a hosted decision model. Plot a reliability curve for each, and record the threshold that gives 95% precision on each model. If the two thresholds differ meaningfully, you have measured the portability gap yourself.

## Reader takeaway

Before you rely on Decider for Mac or any hosted Decisions endpoint, build a labelled evaluation set of 100 to 200 examples from your real data. Run it against the local flash model and the hosted model, record the threshold each needs to hit your target precision, and keep those thresholds per model in config, never hard-coded. Treat "change the URL" as a migration that needs re-calibration, not a configuration change.

## Open question

Will decision probabilities stay vendor-specific scores that each team has to recalibrate, or will vendors start publishing calibrated, comparable probabilities so that a 0.8 from OpenAI and a 0.8 from Perplexity mean the same thing?

## Sources

- [Decider for Mac](https://docs.perplexity.ai/docs/decisions/decider-for-mac)

- [Perplexity releases Decider for Mac to run pplx-decider-flash-v1.2 locally with no API key](https://community.perplexity.ai/t/you-can-now-run-pplx-decider-flash-v1-2-locally-on-your-mac-without-an-api-key/6413)

- [OpenAI developer community announces Decisions API general availability with GPT-Live, caching and 400-question requests](https://community.openai.com/t/decisions-api-is-now-generally-available-with-new-updates/1408934)

- [OpenAI DevDay 2026 Recap for Developers](https://www.infoq.com/news/2026/10/openai-devday-2026)

- [Perplexity's Decider V1.2 models now available on AI Gateway - Vercel](https://vercel.com/changelog/perplexity-decider-v1-2-models-now-available-on-ai-gateway)

- [Ultrafast models, Decisions API, and connector controls](https://www.perplexity.ai/en-GB/changelog/ultrafast-computer-models-decisions-api-and-connector-controls)

- [pplx-decider-v1.2 — Hugging Face](https://huggingface.co/perplexity-ai/pplx-decider-v1.2)

- [pplx-decider-flash-v1.2 — Hugging Face](https://huggingface.co/perplexity-ai/pplx-decider-flash-v1.2)

- [GPT-6 Luna Model — OpenAI API](https://developers.openai.com/api/docs/models/gpt-6-luna)

- [perplexity-ai/pplx-decider-flash-gguf-v1.2 · Hugging Face](https://huggingface.co/perplexity-ai/pplx-decider-flash-gguf-v1.2)

- [pplx-decider-flash-v1.2 — MLX 8-bit · Hugging Face](https://huggingface.co/midium-ai/pplx-decider-flash-v1.2-8bit)

- [Jakevin/pplx-decider-flash-v1.2-GGUF · Hugging Face](https://huggingface.co/Jakevin/pplx-decider-flash-v1.2-GGUF)

Author: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Openai](https://newruntime.com/topics/openai.md): Explore the openai topic hub.
- collection: [Analysis](https://newruntime.com/analysis.json): Open the latest authorial analysis.
- collection: [Source Ledger](https://newruntime.com/source-ledger.json): Inspect the public source evidence graph.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)