OpenAI priced its Decisions API at $0.10 in and nothing out, and Microsoft priced Decision-1 at Jev's $0.042. Neither is the cheapest any more, and OpenAI's has no cache rate.

In our October 5 post on decision models we guessed that if OpenAI shipped Decisions at Luna's $0.10 input with output free, a 356-token ticket would cost about $0.036 per 1,000. It shipped on October 6 at exactly that. $0.0356.

That is still more than twice what Microsoft-Decision-1 and Jev charge, and five times the $0.02 that Perplexity's new Decider costs. The more interesting number is one OpenAI left off: there is no cached-input rate. If your classifier carries a long fixed rubric, the plain Responses API with caching is now cheaper than the endpoint built for the job.

The two new cards

From OpenAI's Decisions guide and changelog, Microsoft's announcement and the Foundry catalog, cross-checked against OpenRouter's live endpoint data on October 14, 2026.

OpenAI's guide puts it plainly: you pay for input tokens and nothing else. Two surcharges carry over from GPT-6 Luna's main card. Regional processing adds 10%, and the guide says long-context multipliers apply without giving a figure for Decisions. Luna's own rate doubles to $0.20 above 272K input tokens, and LLM Gateway lists Decisions the same way. Nobody sends a 272K ticket to a classifier, so we'd ignore it.

Microsoft priced Decision-1 at $0.042, exactly TypeSafe Jev's price. You'll see $0.04 in some write-ups. That's rounding. The 32K window is the thing to check before you switch: Jev takes 64K per request and Perplexity's Decider 262K.

Meanwhile the rest of the market moved

Four rows from our October 5 table are already out of date. On October 9 Cloudflare cut Clef-flash from $0.09 to $0.038 and shrank its context from 64K to 24K, the same day it launched Clef-omni at $0.15 for audio and video states. On October 7 Perplexity replaced pplx-decider v1 ($0.04) with v1.1 at $0.02. And OpenRouter now lists Liquid's d1 at $0.02 from Liquid itself, while Liquid's launch post still says $0.04. We can't tell you which one is current, so we show both below. Mercury Decide, free tier only on October 5, now lists at $0.02, though OpenRouter flags that endpoint as discounted, so it may not stay there.

Output is free on every row. OpenRouter also carries a PrimeIntellect endpoint for Clef ($0.042) and Clef-flash ($0.021) with 16K context; those aren't Cloudflare's prices.

One ticket, nine prices

Same test as last time: eesel's 356-token support ticket with three questions (queue, priority, safe to auto-reply). Decision models bill the input only. The two Luna chat rows add 23 output tokens at $0.50, except the medium-reasoning row, which is eesel's measured bill.

• GPT-6 Luna, medium reasoning$0.0890
• Cloudflare Clef$0.0854
• Cloudflare Clef-omni$0.0534
• GPT-6 Luna, Responses API$0.0471
• GPT-6 Luna Decisions$0.0356
• Microsoft-Decision-1$0.0150
• TypeSafe Jev 1.13$0.0150
• Cloudflare Clef-flash$0.0135
• pplx-decider v1.1 / d1 (OpenRouter)$0.0071

USD per 1,000 tickets, list prices read October 14, 2026, no caching. Solid accent: the two models this post is about. Faded accent: other decision models. Grey: Luna through the ordinary Responses API.

Decisions takes 24% off Luna on the Responses API. That's the output it no longer charges for, and as we said last time, a classifier doesn't write much. Against Luna with medium reasoning the saving is 60%, which is the fair comparison if you were letting it think.

At ten million tickets a month that's $356 on Decisions, $150 on Decision-1 or Jev, $135 on Clef-flash and $71 on Decider v1.1. Real money for a big support desk, rounding error for most teams.

The cache rate OpenAI didn't print

Plenty of classifiers aren't a short ticket. They're a short item plus a long, fixed rubric: label definitions, a few examples, edge-case rules. On the Responses API that rubric caches and bills at Luna's $0.01. On Decisions it bills at $0.10 every time. (OpenRouter shows a cache-read price of 0 for Decisions, but that means no cache line exists, not that cached tokens are free.) Beri ran this case first; here it is with our 23-token output assumption, for a 2,000-token rubric and a 300-token item.

Cached row: 2,000 tokens at $0.01, 300 at $0.10, 23 output at $0.50, cache assumed warm (the first write is $0.125). Decision model rows bill all 2,300 input tokens at list.

The break-even is simple. Decisions only beats a cached Responses call when Luna would have written more output tokens than 18% of the cached prefix. For a 2,000-token rubric that's 360 output tokens, which you'd only see with reasoning turned well up. So if your prompt has a big constant part, run the numbers before you move it. We didn't expect the endpoint built for classification to lose to the general one on price.

Per-token price isn't per-request price

Each vendor counts a decision request its own way, because the questions and options get tokenized along with your text.StackToHeap ran 770 BANKING77 rows through Decisions and Jev on October 8. With definitions only, Decisions billed $0.1057 and Jev $0.0690. With two examples per intent, $0.3024 against $0.1831.

Divide by list price and Jev billed about 55% more tokens than Decisions for the same rows, and 44% more with examples. So OpenAI's 2.38x price gap shrank to 1.53x and 1.65x on the invoice. Same lesson as last time: read usage.input_tokens on your own traffic before you trust any per-1,000 figure, including ours.

Decisions was also the faster of the two in that test, 144 ms p50 against Jev's 298 ms with definitions only.

Does Decision-1 earn Jev's price?

Microsoft says it does. Its announcement claims the highest accuracy in a 36-benchmark comparison and P50 latency about 35x faster than GPT-6 Sol. The page text gives no percentages, but eesel reports Microsoft's chart as Decision-1 at 83.5%, Jev 1.13 at 82.3% and Luna Decisions at 79.4%, with Decision-1 at 85 ms p50.

The one independent run we found disagrees. A dev.to author put 82 hand-labelled questions through 14 decision models on October 10. pplx-decider v1.1 got 80 right. d1 and Jev got 79, Luna Decisions 76 and Decision-1 75, sixth place. Decision-1 was steady when the option order flipped (0 of 26 answers changed) but it was the only cloud model that followed an "ignore previous instructions" line planted in the input.

82 questions is a small test, and Microsoft's set is nearly 150,000. We wouldn't crown anyone on it. It does mean the $0.02 model came top and the $0.10 one came near the bottom, which is not the order the prices suggest. If you feed user text into a router, test injection on Decision-1 before you ship it.

Where we'd put our traffic

For short-input routing and triage, pplx-decider v1.1. It's the cheapest hosted card at $0.02, topped the only independent test we have, has a 262K window and ships Apache 2.0 weights. Keep Jev running beside it as the control.

Luna Decisions makes sense if you're already deep in OpenAI for zero data retention, EU residency or HIPAA. Image states aren't a reason on their own: Decider v1.1, Mercury Decide and Clef-flash take images too, and only Decision-1, Jev and d1 are text-only. Just don't move a long cached rubric onto it. One more thing from Hacker News: a tester saw an option's probability drop from 86% to 73% just by moving it from first to last in the list. Shuffle your options in testing.

Decision-1 we'd wait on. Same price as Jev, half its context, a preview label in Foundry and one bad injection result. Microsoft says it will rebase it on its own and OpenAI models later, and that version may be a different conversation. If you want the generative side of this, Claude Haiku 5.5 now matches Luna's $0.10 and $0.50 below 100K tokens.

What we read for this

• OpenAI: Decisions guide, changelog and pricing - $0.10 input only, no cache or output charges, beta date, surcharges
• Microsoft: Decision-1 announcement and Foundry catalog - $0.042 input, Qwen3.5-9B base, 32,768 context, preview status, speed claims
• eesel: Microsoft-Decision-1 pricing - Microsoft's chart figures; and the ticket test we reuse
• Cloudflare: Clef-omni and Clef-flash changes - Clef-flash $0.038, 24K context, Clef-omni $0.15
• Liquid AI: d1 - $0.04 on the launch post; OpenRouter's endpoint data shows $0.02
• Beri: Decisions API and the cache discount - the cached-rubric case
• StackToHeap: Jev vs Decisions on BANKING77 - bills, accuracy, latency
• dev.to: Decision-1 vs 14 decision models - 82-question test, order flips, injection
• Hacker News: Decisions API thread - option-order bias and latency
• OpenRouter endpoint data - live prices and context for every decision model, read October 14, 2026