Gemini Decisions API pricing (2026): per-token cost, media math, and rate-limit tiers

TL;DR

Gemini Decisions API pricing is simple: $0.10 per million input tokens, and $0 for output because the model never writes any. A 600-token ticket with three questions costs about $0.00008, and a 10-minute call recording about $0.0015. The rate matches the OpenAI Decisions API and is roughly 2.4x TypeSafe Jev per token, so you're paying extra for audio, video and PDF input. The real costs sit elsewhere: no caching, no batch, a training-data free tier, and a Tier 1 cap of 50,000 requests a day. For support teams, triage is now nearly free; writing the answer is still the expensive part, which is the job eesel's AI helpdesk teammate does. With the eesel CLI, a coding agent can upload your routing rules and macros, then run test tickets before anything goes live.

What does the Gemini Decisions API cost?

Gemini Decision 1 (gemini-decision-1) has one price on the paid tier: $0.10 per million input tokens for text, images, video, audio and PDFs alike. Output is listed as "Not applicable (0 output tokens)" on both tiers. That's the whole rate card.

Here's the full row from Google's pricing page, checked November 12, 2026:

The billing rule that makes this cheap: shared input is billed once per request, no matter how many questions you ask. Each extra question only adds its own instruction and criteria tokens. If you've read my explainer on the Gemini Decisions API, you know a request can carry 1 to 1,000 questions over the same input, so ten checks on one ticket cost a little more than one check, not ten times as much.

I've spent two years doing SEO for eesel, and the search behind this page is someone with a spreadsheet open, trying to decide whether a decision model is worth wiring in. So the rest of this post is the spreadsheet: cost per item, the limits that cap your volume, and the costs the rate card leaves out.

How much is one decision, really?

Per-million rates are hard to feel, so here's what the price means per item. The text math is simple. For media, I've used Google's media resolution table for its current models, since Decision 1 accepts the same low, medium, high and ultra_high settings on image and video blocks.

Question tokens are my estimate of about 60 per question. Media figures are approximate; the response's usage.input_tokens_by_modality field tells you the exact count, with video frames and PDF page images counted as image tokens.

Three things fall out of that table:

1. Text triage is a rounding error. Even a big queue costs single-digit dollars a month to classify.
2. Resolution is your biggest lever on images. Dropping a photo from high to low cuts its cost by 75%. For "is there visible damage?" that's usually fine; for reading a serial number, it isn't. Google's own guidance is that PDF quality "typically saturates at medium."
3. Audio is the surprise bargain. A 10-minute voicemail costs $0.0015 to score, with no separate transcription bill. That opens up sentiment or escalation checks on calls that nobody would have paid an LLM to read.

What does the free tier actually give you?

The free tier costs nothing, but the limits and the data terms decide who can use it. Per Google's rate limits page, Gemini Decision gets 200 requests a minute, 50,000 tokens a minute and 10,000 requests a day for free.

That's plenty for building and testing. At 780 tokens a ticket, the token limit lets you run about 64 tickets a minute, and the daily cap covers 10,000. A busy queue could almost run on it.

Don't. On the free tier, Google's pricing table says your content is "Used to improve our products". Customer tickets carry names, order numbers, and sometimes payment or health details. That belongs on the paid tier, where it isn't used for training, alongside the usual checks from your AI guardrails. The paid tier also costs so little at triage volumes that the free tier saves you a few dollars at most.

How do the rate-limit tiers work?

Paid limits rise in steps, based on how much you've spent across Google Cloud and how long ago you first paid. Moving off the free tier means linking a billing account and prepaying at least $5 in credits, per Google's billing docs.

The odd part is the gap between spend and limits. Tier 2 needs $100 of cumulative Google Cloud spend, but at $0.10 per million tokens, $100 is a billion tokens of decisions. If Decision 1 is the only thing you pay Google for, you could sit on Tier 1 for a long time. The 50,000-requests-a-day cap is the one that bites: a queue with 60,000 tickets a day won't fit, however few tokens each one uses.

If you already spend on Google Cloud, you'll likely start higher. Otherwise there's a rate limit increase form, and it's worth filing before launch rather than after your first 429.

How does it compare with other decision models on price?

On price per token, Gemini sits in the middle of the decision models category. Here's each vendor's own published rate, applied to the same 10,000 tickets at 780 tokens each:

Gemini and Flash-Lite figures checked November 12, 2026. Rival rates from each vendor's page as last checked on October 11, 2026. Flash-Lite assumes about 20 output tokens per answer and no thinking tokens.

For plain-text tickets, Jev, Microsoft's model and Clef-flash all cost less than half as much. The difference is pennies at this volume, though, so it only matters at very high scale. Gemini's price makes sense when your inputs aren't text: no other hosted decision model in this table takes audio and PDFs natively, and paying a separate speech-to-text or OCR service first would cost more than the decision itself.

The comparison with Flash-Lite is the one most Gemini users will actually face. Google says Decision 1 is "nearly 5x cheaper" than Flash-Lite on its benchmark suite, and my example comes out at about 4.9x. One Hacker News commenter at launch framed the price trade-off fairly:

"Its bit more expensive than other other Decision Models in the same lane. However, it can accept Audio and Video as input as well and for those specific cases it can be extremely useful."

waximabbax, Hacker News

If you're weighing the rest of Google's lineup, my Gemini 4 Carbon pricing and Gemini 3.8 Flash pricing breakdowns cover the generative side.

Work out your own monthly bill

Plug in your own volumes. The calculator adds up text, photos, calls and PDFs at Google's paid rate, and checks the result against the rate-limit tiers.

What costs aren't on the rate card?

The per-token price is the easy part. These are the things that change what you actually pay, or whether a request goes through at all:

• No caching, ever. Long shared context, like a policy document or a product catalogue sent with every ticket, is billed in full on every request. With generative Gemini models you'd get cached input at a fraction of the price. Here, keep shared context short or move it into code.
• No Batch API or Flex tier. Overnight backfills cost the same as live traffic. On models that support batch, Google's pricing page shows a 50% discount.
• Prepay credits expire. Under Google's prepay plan, unused credits "expire after 12 months" and are non-refundable. When your balance hits $0, every API key on that billing account stops at once with an HTTP 402. Turn on auto-reload, or your triage silently stops routing.
• Overages during billing lag. Google warns that billing data can lag about 10 minutes, so you can run past a project spend cap before it kicks in. At Decision 1's prices that's usually trivial, but it matters if a bug loops requests.
• Safety blocks return errors, not answers. Inputs that trip Google's safety filters come back as HTTP 400. You'll need a fallback path, typically sending the ticket to a person through your normal escalation rules.

Then there's the engineering cost, which is the big one. The API gives you probabilities. You still write the thresholds, the routing logic, the test harness and the code that writes labels back into your helpdesk. The Hacker News thread summed up the use cases as "Routing, classification, replacement for LLM-as-judge", which is right, and each of those is a project, not a toggle.

What does this mean for a support team's budget?

For ticket routing, the model cost has basically gone to zero. That changes where the money in AI support goes, not whether there is any.

Here's a real-world scale check. smava, a German loan comparison site, sends about 27,000 Zendesk tickets a month through eesel across two brands. In one month eesel categorised 14,600 of them, writing language, reply eligibility and category into Zendesk fields, and prepared 15,300 draft replies from 820 synced macros and articles. If you priced only the categorising step on Decision 1 at my 780-token estimate, all 27,000 tickets would come to about $2.11 a month.

The other numbers in that story are where the work is. Drafting 15,300 replies in German that say exactly what an approved macro says, in regulated finance, isn't a decision-model job. Neither is keeping two brands' content apart. That's why I'd frame the Gemini Decisions API as a cheap way to make your triage layer smarter, and not a replacement for the part that answers customers. For the full picture of what AI support spend looks like, my AI customer service cost breakdown covers it, and Zendesk AI pricing shows what a helpdesk vendor charges for the same jobs.

eesel's AI helpdesk teammate does both halves inside your existing queue: it labels and routes tickets in Zendesk, Freshdesk or HubSpot, then drafts or sends the reply from your own help content.

SamCart runs one agent on Freshdesk tickets and another on its website chat, answering from 280 help center articles. Every rollout is simulated against historical tickets first, because a cheap wrong label is still wrong.

Is the Gemini Decisions API worth paying for?

Yes, if you have engineers and your inputs go beyond plain text. At $0.10 per million tokens with free output, the model bill for triage, policy checks or reply QA is close to nothing, and the audio and PDF support saves you a separate transcription or OCR step.

Pay for it if you score calls, photos or documents, you need more than 32K tokens of context, or you run many checks on the same input. Look elsewhere if your tickets are plain text and volume is huge (Jev alternatives and Clef-flash are cheaper per token), you lean on caching or batch discounts, or you need it on Vertex AI today. If none of Google's lineup fits, my Gemini alternatives roundup covers other providers.

And if what you want is triage plus the reply, without building the plumbing, compare the tools in my best AI for ticket triage roundup, or read how to route tickets with intent detection.

Try eesel when the label isn't the hard part

Gemini can tell you a ticket is a billing issue for a fraction of a cent. eesel takes it from there in Zendesk, Freshdesk or HubSpot: it sets the fields, routes it, and drafts the reply your team would have written, tested on your past tickets before it goes live. smava uses it to sort and draft 27,000 German tickets a month. Check eesel pricing, or Try eesel free.