Gemini Decisions API vs OpenAI: caching changes the cost

Google launched the Gemini Decisions API on November 11, 2026. It's a new endpoint, run by the gemini-decision-1 model, that returns probabilities for yes/no, multiple-choice and rubric questions instead of generated text. It costs $0.10 per million input tokens with no output charge, the same uncached rate as the OpenAI Decisions API. For teams already sending routing or policy checks to a model, the choice between the two depends on what the input is and how much of it repeats.

The catch is caching. Google's docs say the Gemini endpoint "doesn't support context caching" and add: "Repeated input across requests counts in full." OpenAI charges $0.01 per million tokens for cached input. If every call carries the same long policy document or reference image, OpenAI can cost much less. If the input is a call recording, a video clip or a PDF, OpenAI's docs list no way to send it directly.

What is the Gemini Decisions API?

It's a scoring endpoint, not a chat endpoint. You send one shared input and a map of questions, and you get one typed answer per question, each with a confidence value from 0 to 1. The Decisions API docs define three question types:

• noul (yes/no) returns the probability that a statement holds, always between 0.01 and 0.99.
• choice picks one of 2 to 255 options you supply and returns a probability for each.
• score rates the input on ordered rubric levels and returns a probability-weighted level, which can fall between levels.

A request can carry up to 1,000 questions over a 262,144-token context window. Google says the model scores options "in a single forward pass" with 150 to 250 ms end-to-end latency, and that, for contexts up to 16K tokens, latency stays "nearly flat from 1 to 100 questions per request." Those are Google's figures, not independent measurements.

The API runs on the Gemini Developer API, v1beta, with SDK 3.2 or later (google-genai for Python, @google/genai for JavaScript). Google says support for Gemini in Google Cloud is "coming soon."

How does it compare with the OpenAI Decisions API?

The two endpoints share a shape: three question types, input-only billing and a probability or confidence on every answer. They differ on inputs, question count, caching and data terms. The table below comes from each vendor's current docs and launch posts.

Scroll sideways to see all 3 columns.

Comparison dimension	Gemini Decisions API	OpenAI Decisions API
Model	gemini-decision-1	gpt-6-luna
Status	v1beta, launched Nov. 11, 2026	Generally available since Oct. 22, 2026
Questions per request	Up to 1,000	Up to 400
Inputs	Text, images, audio, video, PDF	Text, images, JSON, conversation with tool calls
Uncached input	$0.10 per 1M tokens	$0.10 per 1M tokens
Cached input	Not available	$0.01 per 1M tokens
Output tokens	Not billed	Not billed

The input row is the hard boundary. OpenAI's Decisions guide says "files, audio, and item references aren't supported." Gemini accepts audio, video and PDF blocks directly, up to 100 MB of inline data per request (50 MB for PDFs). Google's own example pairs a call recording with account metadata in one request.

The reverse is also true. Gemini's docs say conversation history and tool-call blocks "aren't supported as input blocks," so an agent transcript has to be flattened to text first. OpenAI takes messages, function calls and their outputs as structured input.

When does caching change the bill?

Caching matters when a large, identical block of input repeats across many requests. OpenAI caches the shared input by default once the reusable prefix reaches 1,024 tokens. Gemini bills every input token on every call.

Here's an illustration at list prices. Suppose you check 1 million support tickets against the same 5,000-token policy document, and each ticket adds 200 tokens.

Scroll sideways to see all 3 columns.

1 million requests	Gemini Decisions API	OpenAI Decisions API, cache hit on the document
Policy document (5,000 tokens)	$500	$50
Ticket text (200 tokens)	$20	$20
Total	$520	$70

That's BaristaLabs arithmetic, not a vendor figure. It assumes every OpenAI request hits the cache and leaves out OpenAI's cache-write charges, which its pricing guide lists separately. It also ignores accuracy, which neither vendor's numbers settle for your data.

On Gemini, shared input is billed once per request "no matter how many questions you evaluate." If you ask 300 questions about one document in a single request, you pay for the document once. The expensive pattern on Gemini is the opposite: the same long context sent again for every new item.

Short, unique inputs cost the same on both. A 400-token chat message routed to one of five queues has nothing to cache.

Gemini fits media inputs and long rule lists

BaristaLabs' interpretation: Gemini fits when the input is media or when one item needs a very long checklist.

• Audio and video checks. Call-center QA, voicemail triage or clip moderation, where OpenAI's endpoint would need a transcription step first. Google says there are no "separate audio transcription and OCR costs."
• Scanned documents. Claim forms or invoices as PDFs, scored directly.
• Large rule sets. Checking one input against more than 400 rules in one call.

OpenAI's endpoint fits better when a long reference block repeats across calls, when the input is an agent conversation with tool calls, or when you need zero data retention or HIPAA support. OpenAI says both are available to eligible customers. The Google pages cited here don't mention equivalent terms for this endpoint.

What to check before sending real data

Three details in Google's docs can trip up a first integration.

The free tier uses your content to improve Google's products. The Gemini API pricing page marks free-tier content as "used to improve our products" and paid-tier content as not used. Prototype with synthetic or public examples until the project is on a paid tier. This is the same project-level care we covered when Gemini API began storing requests by default.

Option keys are sorted alphabetically. The API sorts choice keys in UTF-8 byte order, and ties go to the alphabetically first key. Google warns against keys like none_of_the_above, because alphabetical sorting places them ahead of keys like shipping and technical instead of last. Use a plain catch-all such as other with its own description.

Yes/no answers stop at 0.99. A threshold above 0.99 never matches. Set thresholds from labeled examples, as Google recommends, and check what a confidence of 0.9 actually means on your traffic. Our guide to calibrating AI scores before an approval threshold walks through that.

If you call models through Vercel AI Gateway, check the route. Vercel says its TypeSafe-compatible and Decisions APIs support text and image inputs for this model, and PDF, audio and video inputs are supported through AI SDK 7.

Recommended first step

BaristaLabs recommends picking one decision your app already makes at volume and running it on both endpoints before moving traffic.

1. Pull 200 to 500 past items that a person already labeled, such as ticket queues or a pass/fail policy check.
2. Write the same questions for both APIs. Keep option descriptions short and distinct, and include an other option.
3. Record accuracy per confidence band, plus input tokens per request from each response's usage field.
4. Price a month at your real volume, counting how much of each request is repeated context.
5. Keep real customer data out of any free-tier Gemini project.

If your current setup is a chat model with a parsing step, the same labeled set can also show whether a local decision model such as Ollama's d3 would do the job without a hosted call. For help setting up the comparison, see BaristaLabs AI consulting.

Sources and scope

• Google, Gemini Decisions API: Fast, calibrated multimodal decisions, November 11, 2026.
• Google AI Studio, Decisions API documentation and Gemini API pricing, accessed November 12, 2026.
• OpenAI, Build more responsive apps with the Decisions API, October 22, 2026, and the Decisions API guide, accessed November 12, 2026.
• Vercel, Gemini Decision 1 now available on AI Gateway, November 11, 2026.

BaristaLabs did not call either API for this post. Latency, cost and accuracy figures are vendor-reported, and the cost table is illustrative arithmetic at list prices.