Back to blog
AI Tools7 min read

Gemini Decisions API vs OpenAI and Jev on the same 12 tickets

Jev cost 2.3 cents per 1,000 tickets. OpenAI took 156 ms at the median. Compare Gemini Decisions API pricing, file support and accuracy on the same 36 checks.

Moe Lueker
Benchmark results

Jev cost least on these 12 tickets.

Estimated USD per 1,000 tickets

  • Jev 1.14$0.02272
  • Gemini Decision 1$0.03755
  • GPT-6 Luna$0.06032

12 three-question calls per model. Google and OpenAI costs use measured tokens at uncached paid rates. Jev uses OpenRouter charges. Tax and fees excluded. Raw results (CSV)

The Gemini Decisions API costs $0.10 per million input tokens and charges $0 for output. That's the same uncached input price as OpenAI Decisions. Yet on the same support tickets, Gemini cost about 3.8 cents per 1,000, and OpenAI cost 6.0 cents. The two APIs counted different numbers of input tokens. Google pricing · OpenAI Decisions pricing

I'd start with Jev for low-cost text routing, OpenAI for an app where a person is waiting on the answer, and Gemini for a PDF, audio recording or video. The first two picks come from our measured costs and response times. The third comes from the file types Google's endpoint accepts. Gemini input guide

Support ticket test#

A decision API answers a question whose possible answers you define in advance. For a support inbox, that means you can send a message once and ask which team should handle it, whether the customer wants a refund, and whether they've asked for a person. Your app gets values it can act on immediately.

On November 18, 2026, we sent the same 12 synthetic tickets to each provider. We wrote the expected answers first. Each request contained one ticket and all three questions. Here are the results for those 36 requests.

Tested endpointCorrect checksMedian time (range)Model cost per 1,000 tickets
Jev 1.14 via OpenRouter36/36195 ms (161 to 244)$0.02272
Gemini Decision 1 via Google36/36392 ms (287 to 607)$0.03755
GPT-6 Luna via OpenAI36/36156 ms (126 to 571)$0.06032

Each cost is an estimate for 3,000 decisions across 1,000 tickets of this length. Gemini averaged 375.5 input tokens per ticket, Jev 540.9, and OpenAI 603.2. For Gemini we applied the paid rate to its reported tokens. For OpenAI we used its uncached rate. For Jev we used OpenRouter's reported inference charge. The figures exclude tax, gateway funding fees and the cost of running your app. Download the 45-request benchmark CSV · Read the tickets, questions and expected answers

Which decision model was fastest or most accurate?#

OpenAI had the lowest median here. Jev had the tightest observed range across the 12 tickets. We timed the full response from one Mac, one request at a time, rotating the provider order. That includes the network trip. Jev's result includes a gateway hop. If you're choosing between a gateway and a direct connection, this guide to how OpenRouter works explains the difference.

All three matched all 36 labels. These were short, deliberately written examples, so the tie tells us little about messy real inboxes. A broader test gives another view. On November 18, the Artificial Analysis Decision Index v2.1 scored Jev 66.2, Gemini 66.0, and OpenAI 60.2 across tasks including document decisions and tool selection. Those are index points. Jev was fastest in that setup at 92 ms. If latency matters to your product, measure it through the endpoint your users will hit. Artificial Analysis Decision Index

What changes when you ask three questions in one request?#

One word in ticket 10 makes it a useful test for an inbox rule that looks for the word “charge.”

My replacement box went to the old office even though the receipt shows the new address. I need a person at your shipping desk to check where it went, then send it to the right building. No need to reverse the charge.

All three models returned delivery, refund requested = false, and human requested = true. We then sent each question on its own, repeating the ticket each time. We got the same answers in both versions. Here's what happened to the input bill.

ModelTokens in one callTokens across three callsCost saved
Jev 1.145521,16252%
Gemini Decision 138656432%
GPT-6 Luna61571514%

Google bills the shared file or text once in a request. Each question still adds tokens for instructions, choices and formatting. In this example, Jev's model charge went from $0.00004880 to $0.00002318. We ran each version once on this ticket. For a long document with dozens of questions, I'd measure both shapes again before budgeting. Google billing details

Which API fits your input?#

Gemini Decision 1 for original files. Say you need to know whether someone gave verbal consent on a call. Gemini can score the recording directly, including how something was said. It also accepts video, PDFs, images and text. You can try it on the free tier. Paid input costs $0.10 per million tokens, with up to 1,000 questions per request inside a 262,144-token window. If you have a video on YouTube, supply the media file or a supported media URL. A YouTube watch link won't work on this endpoint. Gemini Decisions guide · Google pricing

OpenAI GPT-6 Luna for an app that already sends text, JSON or images to OpenAI. Its Decisions endpoint accepts up to 400 questions per request. One reason to try it on a document-heavy workflow is prompt caching. If you keep reusing the same eligible input prefix across calls, cache hits get a discounted rate. Cache writes carry a charge. Our requests had zero cache reads or writes, so the table shows the full uncached cost. OpenAI Decisions guide · Prompt caching

TypeSafe Jev 1.14 for cheap, repeated text or JSON checks. Jev lists $0.042 per million input tokens and free output. It also supports experimental image input. On TypeSafe's direct endpoint, the default limits are 64K tokens per request and 32K per question. The number of questions can grow as long as both fit. Our test used OpenRouter's System One endpoint. If a job starts with raw audio or video, budget for a separate processing step or try Gemini first. TypeSafe models · Image support · OpenRouter route

If your customer needs a written reply or a verbatim explanation, use a generative model for that step. A decision score can tell your app to send a ticket to billing. A billing reply needs words. I compared cost per task for that kind of model in my Gemini Hydrogen, Flash, Haiku and Luna comparison.

Reproduce the benchmark#

Download the benchmark kit (.zip) and unzip it. It contains the 12 labeled tickets, the Node.js runner and the original results CSV. Open a terminal in that folder. Set GEMINI_API_KEY, OPENAI_API_KEY and OPENROUTER_API_KEY in your environment. These commands call the three live APIs and write receipts locally.

node run-benchmark.mjs smoke
node run-benchmark.mjs suite

The results CSV contains 36 three-question requests and nine single-question requests. We excluded three warm-up calls. All measured calls succeeded on the first attempt. We counted a yes/no probability of 0.5 or higher as yes. Jev resolved to typesafe/jev-1.14-20261014, OpenAI returned gpt-6-luna, and Google's response didn't name a checkpoint beyond the requested gemini-decision-1.

The first change I'd make for your own test is to swap in tickets people routed incorrectly last month. Label those before you run them. A model that catches even one expensive misroute can be worth more than the few cents separating these three on an easy inbox.

Moe Lueker

Moe shares tool walkthroughs and lessons from real projects. Mechanical engineer, then venture capital, now building AI tools for creators and small businesses. More about Moe

Gemini Decisions APIOpenAI Decisions APITypeSafe JevAI benchmarks