Gemini Decision 1 vs GPT-6 Luna vs Jev 1.14: Decision Models Compared — No Winner Crowned

Google has entered the decision-model race. Gemini Decision 1, launched on 11 November 2026 by Google DeepMind, scores up to 1,000 yes/no, choice and rubric questions in a single forward pass — and Google's own benchmarks put it ahead of OpenAI's GPT-6 Luna and TypeSafe's Jev 1.14. But an independent index tells a different story, ranking Decision 1 ninth out of eighteen decision models. This comparison puts all three side by side on price, speed, accuracy and what they actually are — and crowns no winner, because the numbers point in different directions depending on who is counting.

First, What Is a Decision Model?

A "decision model" is a specialised AI that does not chat. Instead of generating text, it takes a question with a fixed set of possible answers — yes or no, option A or B, a score on a rubric — and returns probabilities and confidence scores for each answer. No paragraphs, no explanation, just numbers. That sounds limited, but it is exactly what a huge amount of real software needs: is this transaction fraud, which of these candidates fits the job, does this document meet the policy, how confident are we in this diagnosis?

Gemini Decision 1: Google's New Specialist

Gemini Decision 1 — model name gemini-decision-1 — launched on 11 November 2026. It takes one multimodal input — text, images, audio, video or PDF — and scores up to 1,000 questions against it in a single forward pass. A "forward pass" is one trip through the model: instead of asking the model one question at a time, you hand it a thousand questions at once and get a thousand scored answers back. The context window is up to 256K tokens (262,144 per the documentation), which means the input document or media can be very large.

AI-generated illustration by INORYVA.

GPT-6 Luna: OpenAI's Efficient Generalist

GPT-6 Luna — model name gpt-6-luna — is a real, released OpenAI model, launched on 22 September 2026. It appears on OpenAI's own pricing page at $0.10 per 1 million input tokens and $0.50 per 1 million output tokens, with cached input at $0.01. Its context window is 1.05 million tokens, and its knowledge cutoff is 18 May 2026. OpenAI describes it as its "most efficient model for focused, high-volume tasks" — the cheap, fast workhorse of the GPT-6 family.

TypeSafe Jev 1.14: The Cheapest Specialist

TypeSafe's Jev 1.14 — version jev-1.14.0 — was released on 14 October 2026. It is TypeSafe's flagship "System One" decision model, and it is the cheapest of the three by a wide margin: $0.042 per 1 million input tokens, or $42 per billion tokens, with output free. That is less than half the input price of either Google's or OpenAI's offering, which makes it the obvious choice on cost alone — if cost were the only thing that mattered.

Jev takes text plus PNG and JPEG images, and returns choices, scores and yes/no answers with probabilities — the same shape of output as Decision 1. Its context is 64K tokens per request, or 32K per question, rising to 262K with an opt-in header. That is smaller than Decision 1's 256K and far smaller than Luna's 1.05 million tokens, which limits how much material you can score in one call.

	Gemini Decision 1	GPT-6 Luna	TypeSafe Jev 1.14
Launched	11 Nov 2026	22 Sep 2026	14 Oct 2026
Type	Decision-only	General-purpose (via Decisions API)	Decision-only
Input price / 1M tokens	$0.10	$0.10	$0.042
Output price / 1M tokens	$0 (no text output)	$0.50	$0 (no text output)
Context window	256K tokens	1.05M tokens	64K per request (262K opt-in)
Max questions per call	1,000	Via Decisions API	Not stated as a single-pass limit
Latency	150–250 ms end to end	Not stated for decisions	Not stated
Availability	Gemini API (AI Studio v1beta); Vertex AI coming soon	OpenAI API; Decisions API GA 22 Oct 2026	TypeSafe API

The Benchmarks: Google's Numbers vs the Independent Verdict

Google's own benchmark table is the centrepiece of its launch, and it should be read as exactly that: Google's own. On JevBench — a benchmark named after TypeSafe's model, which Google chose to run its comparison on — Decision 1 scored 92.2%, against 90.9% for GPT-6 Luna and 90.0% for Jev 1.14. Decision 1 led five of the seven tests all four models ran, and seven of nine tests overall. On calibration — how well a model's confidence matches its actual accuracy, measured as expected calibration error — Decision 1 scored 0.0250, against 0.0721 for Luna and 0.0494 for Jev. Lower is better, so Decision 1's confidence is the best matched to reality of the three. Google also claims Decision 1 is roughly five times cheaper than Gemini 3.5 Flash-Lite with the Interactions API across the shared test suite.

The independent check exists, and it disagrees. Artificial Analysis, an independent benchmarking organisation, runs a Decision Index, and version 2.1 puts Gemini Decision 1 at 65.98 — ninth out of eighteen decision models. Not first. Not top three. Ninth. That is not a contradiction of Google's table so much as a different measurement: a different set of tasks, a different scoring method, a different field of competitors. But it is a warning against crowning a winner on the basis of one vendor's benchmark. On JevBench, Decision 1 leads. On the Decision Index, it sits in the middle of the pack. Both can be true at once — and that is precisely why this article crowns no winner.

AI-generated illustration by INORYVA.

The Apples-to-Apples Warning

There is a deeper problem with this comparison, and it is the one that matters most. Decision 1 and Jev are decision-only models. They output scores and probabilities, never text. GPT-6 Luna is a general-purpose model that is used as a decision model through OpenAI's Decisions API. Comparing them is only valid as a comparison of decision APIs — of the three ways to get a scored yes/no answer out of an API call — and not as a comparison of the models themselves.

Which One Should You Pick?

AI-generated illustration by INORYVA.

What to Watch Next

Three things will settle this race. First, Vertex AI support for Decision 1: until it arrives, Google's enterprise customers cannot use the model through their normal channel, and that limits its reach. Second, independent benchmarks: the Artificial Analysis Decision Index is one data point, and more independent tests will show whether Google's JevBench lead holds up outside Google's own test suite. Third, TypeSafe's "danger zone": if the image, Box and long-context features graduate to production-ready, Jev's price advantage becomes much harder to ignore.

Key Takeaways

• Gemini Decision 1 (11 Nov 2026) scores up to 1,000 questions in one forward pass, at $0.10 per 1M input tokens and $0 output, with 150–250 ms latency and a 256K-token window.
• GPT-6 Luna (22 Sep 2026) is a general-purpose model at $0.10 input / $0.50 output per 1M tokens, with a 1.05M-token window — a decision model only via OpenAI's Decisions API (GA 22 Oct 2026).
• TypeSafe Jev 1.14 (14 Oct 2026) is cheapest at $0.042 per 1M input tokens, output free — but its image, Box and long-context features are in TypeSafe's "danger zone", not production-ready.
• On Google's own JevBench table: Decision 1 92.2% vs Luna 90.9% vs Jev 1.14 90.0%; calibration error 0.0250 vs 0.0721 vs 0.0494.
• The independent Artificial Analysis Decision Index v2.1 ranks Decision 1 at 65.98 — 9th of 18 decision models, not #1 overall.
• The comparison is only valid as a decision-API comparison: Decision 1 and Jev output scores only, while Luna is a general-purpose model. No winner is crowned.
• Do not conflate Luna with GPT-6.2, which has not been released as of 12 Nov 2026.

Conclusion

Three models, three strengths, and no winner. Decision 1 is the fastest specialist for bulk scoring, Luna is the most flexible because it is a general-purpose model behind a decisions API, and Jev is the cheapest by a wide margin — with a warning label on its newest features. Google's benchmarks favour its own model; the independent index does not. The right choice is the one that fits your workload, your budget and your tolerance for preview features, measured on your own data. Run the test yourself. That is the only benchmark that counts.

Sources

Google — Gemini Decisions API (Gemini Decision 1)

TypeSafe — Models (Jev 1.14)

VentureBeat — Google's Gemini Decision 1 Model Scores Up to 1,000 Questions in One API Call