In partnership with
A new kind of model got its own lane in October, and none of them writes a sentence. OpenAI opened its Decisions API on Oct 6. Liquid AI released d1-ultra on Oct 15. Per Vercel data, TypeSafe's Jev reached 35% of Vercel AI Gateway paid teams within two weeks of its limited release.
I think a lot of apps are paying a generating model to say "yes." Look at your own code. If a call's output gets parsed straight into a label, a route, or a score, you're paying for output tokens your code reduces to one value. Decision models answer those questions with probabilities and bill input tokens only. It's a narrow job, and for that job they fit.
The Signal
A decision model takes your input plus a question with a fixed answer set and returns probabilities. Ask which of your five queues a ticket belongs in and you get one of your five values, with probabilities attached. None of the three below charges for output tokens.
OpenAI's Decisions API went into public beta on Oct 6. The endpoint is POST /v1/decisions, and the only model is gpt-6-luna. You send text, images, or both, plus three question types: predicate (probability from 0 to 1 that a condition is true), choice (one of your supplied values, with probabilities), and score (a probability-weighted rating against ordered levels). OpenAI says it returns typed answers "about 10x faster than the Responses API" and expects general availability "in the coming weeks."
The price is $0.10 per million input tokens, with no charge for output, cache reads, or cache writes. Regional and long-context multipliers still apply. Through the Responses API, the same model is $0.10 input and $0.50 output per million, so the output charge disappears on this endpoint. ZDR and HIPAA are available for eligible customers. (OpenAI docs, OpenAI changelog)
Liquid AI released d1-ultra on Oct 15. It reads text and images and returns probabilities for yes/no, choice, and score questions "in one forward pass," without generating tokens. It costs $0.042 per million input tokens, input only, with a 32,768-token context. Each question is billed as its own prompt, including its text and images, and a 1024x1024 image counts as 1,536 tokens. Four questions about one screenshot means paying for that screenshot four times. Liquid says d1-ultra outperforms Jev 1.14 across its 112 evaluations at the same price. That's a vendor claim, so test it. The model ID on Vercel AI Gateway is liquid/d1-ultra, and it's also on OpenRouter. (Liquid AI, Vercel)
TypeSafe AI introduced Jev in a limited release on Sept 15. Per Vercel data reported by TechCrunch on Oct 16, it was used by 35% of Vercel AI Gateway paid teams within two weeks, nearly double the reach of the second most popular model, Sonnet 5.5. Vercel's report also put Jev at 0.27% of gateway spend over that same period. Vercel itself replaced GPT-6 Luna with Jev for classifying harmful commands.
Jev is text only, $0.042 per million input tokens, output free, at POST /v1/systemone. Current versions are jev-1.14.0 (since Oct 14) and jev-1.13.0. jev-latest points to 1.13 and jev-preview to 1.14. (TechCrunch, TypeSafe docs, AI Weekly)
Decision models have limits. They don't write, they need a fixed answer set, and their probabilities need thresholds you validate on your own data.
The AI notetaker that gets the hard words right
Your AI is only as good as what you feed it. Feed it a transcript with your product names, acronyms, and numbers spelled wrong, and you spend the afternoon hand-fixing the output.
Wispr Flow Notetaker uses your dictionary and calendar, so product names, acronyms, numbers, and uncommon names come out spelled right. It pops up when your meeting starts and captures Zoom, Google Meet, Teams, and Slack huddles with one click, no bot joining the call.
Then your meetings show up inside Claude or ChatGPT through the built-in connector. No copy paste. Try Notetaker free on Mac and Windows.
The Stack
- Search your app for model calls whose output you immediately parse into a label, boolean, route, or number. Think ticket routing, moderation, lead scoring, and "should the agent retry, stop, or ask."
- For each, write the fixed answer set and the threshold that triggers an action.
- Run 50 to 100 real past inputs through one decision model (
gpt-6-lunaon/v1/decisions,liquid/d1-ultraon a gateway, or Jev) and compare with what your current model returned. - Set the action threshold on that data. Send low-probability cases to a human or to your current generating model.
- If you pin, pin a versioned ID like
jev-1.14.0. Aliases likejev-latestmove when a new release ships. - Keep prose and code generation where it is.
One decision: move one classification call this week. If none of your calls ends in a fixed label, move none.
Prompt of the Day
Paste this with a list or code excerpt of your app's model calls:
Below are the model calls in my app. For each one I've given the call name, what it sends to the model, and what my code does with the output.
Mark each call "decision" or "generation." A call is a decision if my code parses the output into a fixed label, a boolean, a route, or a number. A call is generation if the output is shown to a user or used as prose or code.
For every decision call, write the full answer set and suggest a question type: predicate (the probability that one condition is true), choice (one value from a fixed list), or score (a rating against ordered levels).
Use only what I give you. Do not estimate cost, volume, or accuracy numbers I didn't provide. If you can't tell what my code does with an output, write "unclear" and move on.
If it marks a call as a decision and can't write the answer set, throw that row out.
Bring your sorted list to AI Freedom Lab, where builders launch companies and web and mobile apps with AI.
Tool of the Day
Replit, featured in the AI Tools Index.
Replit added Jev to Replit AI Integrations on Oct 2. The changelog pitch reads "Build apps that classify content, route requests, and score leads... without managing API keys." You ask Agent to use jev-latest through OpenRouter. TypeSafe's docs point that alias to 1.13. On Oct 16 Replit also shipped Bring Your Own Cloud (Full BYOC in private preview) and put Claude Haiku 5.5 in the model selector for Core and Pro in Power Mode. (Oct 2 changelog, Oct 16 changelog)
Open an existing project today and ask Agent to wire one routing or scoring step to Jev.
Quick Hits
Gemini API's deep-research-pro-preview-12-2025 agent shuts down today, Oct 23. Change the agent parameter in interactions.create to deep-research-preview-04-2026 for speed and streaming, or deep-research-max-preview-04-2026 for maximum comprehensiveness. The same Oct 8 note deprecated gemini-3.7-flash and gemini-3.5-flash, which now auto-route to gemini-3.8-flash and gemini-3.6-flash. (Gemini API changelog)
OpenAI added 12-hour prompt caching to the Responses API on Oct 20. Shared instructions, tools, and context stay cached for requests hours apart. 12-hour cache writes cost 2.5x the input token price, so check that your cached prefix actually gets reused before you rely on it. (OpenAI changelog)
Claude Haiku 5.5 (claude-haiku-5-5) launched Oct 7 with 1M context, 128k max output, and adaptive thinking with effort. Code written for Haiku 4.5 can break: budget_tokens manual thinking, temperature/top_p/top_k, and assistant prefill can each return a 400, and adaptive thinking is on by default. Same day, Sonnet 5.5 cache reads dropped from $0.20 to $0.10 per million tokens. (Claude release notes)
Closing the loop
Every call that ends in a label is a question with a short list of answers. Pick the one that runs most often, run 50 past inputs through a decision model, and set the threshold on what you see. If nothing in your app ends in a label, leave it alone and keep building.
Join AI Freedom Lab, where builders launch companies and web and mobile apps with AI: skool.com/aifreedomlab