Gemini Decision 1 vs GPT-6 Luna: Benchmarks, Cost, Access

Two decision models landed in the same few weeks, and I spent launch day reading both launch posts, both sets of API docs, and the independent Artificial Analysis Decision Index so you don't have to.

The surprise with Gemini Decision 1 vs GPT-6 Luna is that price barely separates them. Both bill $0.10 per million input tokens with zero output cost, so the real fight is over accuracy, calibration, what each one can read, and whether you can even call it from your stack today.

If you want the background on what Gemini Decision 1 actually is, I wrote a separate explainer; for Google's generative models, start with Gemini 4 Argon and the wider Claude vs Gemini comparison. This piece is the head-to-head.

Gemini Decision 1 vs GPT-6 Luna: the short answer

Gemini Decision 1 is the more accurate and better-calibrated model of the two. It leads most of Google's own benchmark rows, and on the independent Artificial Analysis Decision Index it scores 66 to Luna's 60, with the lower Brier score (better calibration) of the pair.

GPT-6 Luna, part of OpenAI's GPT-6 lineup, is the more flexible one to run. It matches that $0.10 list price, adds prompt caching at $0.01 per million cached tokens, speaks to OpenAI's live voice stack, and carries zero data retention (ZDR) and HIPAA options. It just reads less: text, images, and JSON only.

There's a third name you shouldn't ignore. TypeSafe's Jev 1.14 ties Gemini on that same independent index (66), runs faster, and costs less than half per task. On launch day, "which is best" has three honest answers.

Choose Gemini Decision 1 if you:

• Need the top accuracy and the tightest calibration of the two frontier options.
• Feed it mixed media: call recordings, scanned PDFs, video frames, or images alongside text.
• Want the widest batch of questions per call (up to 1,000).

Choose GPT-6 Luna if you:

• Reuse the same context across many calls and want caching to cut the bill.
• Live in OpenAI's world already, especially for real-time voice.
• Need ZDR or HIPAA handling, or inference residency in Europe (EEA plus Switzerland), on day one.

Stick with Jev 1.14 if you:

• Care most about latency and cost per task.
• Can trade a touch of calibration (Brier 0.19 vs 0.18) for a model that's less than half the price per job.
• Already run TypeSafe tooling.

My take: for most teams building a classifier, router, or grader, Gemini Decision 1 is the one I'd prototype first. If your workload leans on reused context or voice, test Luna beside it before you commit.

Specs, pricing, and access side by side

The two models read almost identical on the price line and diverge everywhere else. How they compare at a glance:

Two notes before you build on this table.

The list price is a tie, so per-call cost comes down to tokens and caching. Both charge $0.10 per million input tokens and bill the shared input once per request. If you want the wider picture on how Google meters its models, the Gemini pricing breakdown covers the rest of the family.

I left Luna's context window qualified on purpose. OpenAI publishes 1,050,000 tokens for the base gpt-6-luna model, but the Decisions guide itself does not state a window for that endpoint. If that number matters to your design, test it against your real payloads.

What the benchmarks say: Google's own table

Google's launch post is confident. By its own numbers, Gemini Decision 1 "leads across the suite" with the lowest calibration error among the models Google chose to show. Treat that the way you'd treat any vendor's own chart: it's the best case, measured by the team shipping the product.

Google's full launch-day table, as Google reported it (Jev 1.14 row kept in):

Google's footnote: refusals, safety blocks, and requests a model could not accept are scored as incorrect; every score uses the full benchmark as the denominator. Scores are Top-1 accuracy unless otherwise noted.

Three rows worth pulling out, all as Google reported them:

• Cost (seven shared evals): Gemini Decision 1 comes in around $0.88, GPT-6 Luna around $0.93, and Gemini 3.5 Flash-Lite around $4.14. That last gap is where Google's "nearly 5x cheaper than Flash-Lite through the Interactions API" line comes from.
• Where Jev 1.14 wins: Google keeps TypeSafe's Jev 1.14 in the table, and it takes Mind2Web and When2Call, plus the cost row. Google didn't hide that.
• Latency: Google claims 150 to 250 ms end to end, with 50 to 100 ms of actual model compute, and says it stays nearly flat from one question to a hundred because it scores options in a single pass instead of writing tokens one at a time.

So Google's own framing already admits Jev is cheaper and wins two evals. Hold onto that, because the independent numbers tell a slightly different story than "leads across the suite."

‍

{{templates}}

‍

What independent testing says: Artificial Analysis

When I pulled the Artificial Analysis Decision Index on launch day, both models were already scored, which is rare for a day-one release. Artificial Analysis runs every model through the same setup, so this is the closest thing to an apples-to-apples decision model comparison. Artificial Analysis Decision Index, launch-day scores:

Read it plainly. Gemini Decision 1 beats Luna on the index and on accuracy, and it's the best-calibrated of the three. That last part backs up Google's calibration claim, at least within this trio.

But the gap isn't a blowout. Luna is a touch faster and a few dollars cheaper per task on these runs, and both frontier models sit mid-pack: the index leaders (Drex Max v1.1, SubQ Snap 1.2, and GLiDE v1.2) are tied at 73.

And Jev 1.14 is the quiet story. It ties Gemini on the index (66), and shares that 66 with Elephant 9B. It runs more than twice as fast (92 ms), and costs less than half per task ($156). If your decisions are high-volume and latency-sensitive, that's hard to wave off.

Why Google's table and the independent numbers disagree

Google says "leads the suite." Artificial Analysis says Gemini ties Jev and edges Luna. Both can be right, for the same three reasons that split vendor charts from independent runs in the earlier Gemini 4 Argon vs GPT-6 Astra and Gemini 4 Carbon vs Claude Opus 5.5 comparisons.

The three reasons the numbers split:

• Different test lists. Google's table leans on the evals where Decision 1 looks strongest. Artificial Analysis averages its own six, including ones where other models lead.
• Different sources. Google mixes its own runs with outside leaderboards. Artificial Analysis runs everything itself on one setup.
• Different cost math. Google's cost row is the total for seven shared evals. Artificial Analysis measures cost per task at real token usage, which is why its cost order can flip even when the sticker price is identical.

The honest takeaway: Gemini Decision 1 is a strong, well-calibrated model, but on the independent index it sits mid-pack, seven points behind the leaders.

Gemini Decisions API vs OpenAI Decisions: access and what each cannot do

Gemini Decision 1 is still in v1beta with no Google Cloud support yet, while GPT-6 Luna's Decisions API is already generally available.

Gemini Decision 1 access and limits

Gemini Decision 1 is available on the Gemini Developer API (v1beta), and it needs SDK 3.2 or later. There's a free tier in Google AI Studio for prototyping.

It generates zero text, so there's no text generation, no function calling, no Live API, and no "thinking" mode.

It also has no context caching, no Batch API, and no Flex or Priority inference, which means repeated context counts in full every time. And it isn't on Google Cloud or Vertex yet; Google lists that as coming soon, so re-check before you plan around it.

GPT-6 Luna Decisions access and limits

Luna's Decisions API is generally available through a dedicated /v1/decisions endpoint, and gpt-6-luna is the only model that powers it today. You can ask up to 400 questions per request.

Its edge is the surrounding stack. Luna offers prompt caching at $0.01 per million cached tokens, a native tie-in to OpenAI's live voice product, inference residency in the US and Europe (EEA plus Switzerland), and ZDR plus HIPAA handling for eligible customers. The catch: it reads text, images, and JSON only. OpenAI's own docs say files, audio, and item references aren't supported, so no audio, video, or PDF.

Gemini Decision 1 vs Flash-Lite: why it isn't a fair match

You'll see Flash-Lite in Google's cost row at around $4.14 against roughly $0.88, and it's tempting to call it the budget option. It isn't a fair match: Gemini 3.5 Flash-Lite is a generative model you'd call through the Interactions API, the same family as Gemini 3.8 Flash, so it writes tokens and costs more per decision.

Gemini Decision 1 cost vs GPT-6 Luna: which should you use?

You should use Gemini Decision 1 for accuracy and mixed media, GPT-6 Luna for cached or regulated workloads, and Jev 1.14 when cost per task leads. Pick by the job:

• Go Gemini Decision 1 for the best accuracy and calibration, and for anything multimodal: call audio, scanned PDFs, or video frames.
• Go GPT-6 Luna for reused-context workloads (caching earns its keep), for voice, and for regulated data that needs ZDR, HIPAA, or Europe (EEA plus Switzerland) residency.
• Go Jev 1.14 when latency and cost per task lead the decision.

Here's the money math, using only numbers both vendors publish. Say you send 1,000 requests, each with 5,000 shared input tokens. That's 5 million input tokens, which costs $0.50 on either model at $0.10 per million. Identical.

The tiebreak is caching. If most of that input repeats across calls, Luna's $0.01-per-million cached rate can cut the repeated portion by about 10x. Gemini Decision 1 has no caching, so the same tokens bill in full every time.

With 4 million of those 5 million tokens cached, Luna comes to $0.14 ($0.04 for the cached tokens plus $0.10 for the rest), while Gemini Decision 1 stays at $0.50.

One caveat: the Artificial Analysis per-task costs above ($345 vs $339 per million tasks) differ from this simple math because they reflect how many tokens each model actually burns per task, not the list price.

Pick the model, then put it to work

Choosing a decision model is the easy part. A model that returns a probability, a choice, or a score is a routing signal. It tells your code which lane to take.

The routing call fires in milliseconds, and then a human still has to triage the email, update the record, and send the follow-up.

Lindy is the AI teammate that handles that part. It's model-agnostic, so it works with whichever model you pick.

With Lindy, teams can:

• Triage, label, and route incoming email automatically.
• Schedule meetings and ask for your okay before anything goes out.
• Take meeting notes and draft the follow-ups for you to approve.
• Update the CRM after calls and emails.
• Answer teammates' questions from your knowledge base right in Slack.
• Run lead gen, enrichment, and outreach without a separate tool for each step.

Say a customer email lands in the billing lane. Lindy labels the thread, updates the CRM record, and drafts the follow-up, then sends it once you approve.

Whichever side of Gemini Decision 1 vs GPT-6 Luna wins your tests, route each job to that model and let Lindy close the loop on the work that follows. It's closer to a real hire than another app to configure.

Try Lindy free and see how much of your daily workload an assistant can take off your plate.

‍

{{cta}}

‍

Frequently asked questions

GPT-6 Luna vs Gemini Decision 1: which is more accurate?

Gemini Decision 1 is more accurate than GPT-6 Luna. It leads most of Google's own benchmark rows and scores 66 to Luna's 60 on the independent Artificial Analysis Decision Index, with better calibration (a lower Brier score).

Which is cheaper, Gemini Decision 1 or GPT-6 Luna?

Both cost the same $0.10 per million input tokens with no output charge, so list price is a tie. GPT-6 Luna adds cached input at $0.01 per million, so workloads that reuse context can run cheaper on Luna. TypeSafe Jev 1.14 is far cheaper per task than either.

Can Gemini Decision 1 replace GPT-6 Luna?

Gemini Decision 1 can replace GPT-6 Luna for many routing, grading, and classification jobs, especially multimodal ones, since it reads audio, video, and PDFs. It's a weaker swap if you need prompt caching, OpenAI's live voice integration, or OpenAI's compliance options today.

Gemini Decision 1 vs Jev 1.14: when is Jev the better pick?

TypeSafe Jev 1.14 shares the 66 Artificial Analysis Decision Index score with Gemini Decision 1 and Elephant 9B, runs faster (about 92 ms per task), and costs roughly half per task. It's the value pick when latency and budget lead your decision.

Is Gemini Decision 1 on Google Cloud?

Not yet. Gemini Decision 1 runs on the Gemini Developer API (v1beta), and Google lists Google Cloud and Vertex support as coming soon. Re-check Google's docs before you plan around it.