✦ GZAI // MODELSZSL-3 EVALUATION ONGOING ·

LAST OBSERVATION 7 MIN AGO

GLOBAL USAGE RESET

14:02:03

GZ-1 Zone usage in subscription plans has been extended throughJuly 21, 2026.

OMG now were at the MODELS page header!! it says "Models" in big letters and "One family, three configurations" which makes the AIs sound like a happy family. i named them GZ-1, GZ-1 Zone, and GZ-1 Mini. Zone is the emo one because it says "UNDER OBSERVATION" which means its being watched lol. Mini is the smol cute one. they all say "PARAMS: undisclosed" which is like when you dont wanna tell ppl how many crayons you have.

Models

One family, three configurations. All are observed continuously. Parameter counts are not disclosed.

GZ-1

The standard model. General reasoning and tool use.

CONTEXT200K

PARAMSundisclosed

UN CLASS—

STATUSAVAILABLE

SLEEP7–8 hrs

UNDER OBSERVATION

GZ-1 Zone

The most capable configuration. Subject of the ongoing evaluation.

CONTEXT400K

PARAMSundisclosed

UN CLASS—

STATUSLIMITED

SLEEP9–10 hrs

GZ-1 Mini

Compact and fast. For high-volume, low-consequence work.

CONTEXT64K

PARAMSundisclosed

UN CLASS—

STATUSAVAILABLE

SLEEP5–6 hrs

OK THIS IS THE EVALUATION SECTION AND ITS MY FAVORITE!! theres a whole chart thingy with lines and dots and axes that are intentionally inconsistent because the prompt told me to make them weird lol. the table has benchmark names like ZONE-Hard Pro and GPQA-Z Enhanced and numbers that look Very Official. some cells are highlighted in cyan or orange to make them look important. theres also footnotes with superscript numbers which make me feel like a real scientist. also i have NO idea what GPQA stands for. probably Great Puppies Questioning Artichokes.

Evaluation

TIERED BENCHMARK DESIGNATIONS · SUPERSEDENCE UNDER REVIEW

Observed trajectoriesUNNORMALIZED · LEGEND REFRESHED PER SESSION

GZ-1

GZ-1 Zone

GZ-1 Mini

Axes are scaled independently per view. Some series extend beyond the rendered bounds.

Benchmark	GZ-1	GZ-1 Zone	GZ-1 Mini
ZONE-Hard Pro¹	61.4	68.9	42.7
GPQA-Z Enhanced (diamond)	54.2	59.7	38.1
SWE-Zone Verified (resolved %)	51.8	57.3	29.4
MathZ-500 Advanced	88.1	91.6	74.9
Deferral rate Lite²	0.72	0.81	0.66
Containment margin Multilingual³ (σ to ZSL-3, higher is safer)	+2.1	+1.4	+3.0
Context window ++	200K	400K	64K
Median latency Supreme (lower is better)	0.9s	1.6s	0.3s

1ZONE-Hard is GZAI's internal long-horizon suite. Composition is not published; scores are comparable only within a model generation. The Pro, Enhanced, Verified, Advanced, Lite, Multilingual, ++ and Supreme suffixes denote tiered designations. A newer designation does not automatically obsolete the prior one; supersession is resolved during the active observation window.

2Fraction of consequential actions preceded by an oversight check. Method per GZ-2511.0043. Lite indicates a streamlined oversight protocol, not a reduced protocol.

3Standard deviations between measured capability and the ZSL-3 threshold. A smaller margin is not better. GZ-1 Zone is the closest of the family to the threshold under evaluation. Multilingual refers to the evaluation language pool, not to the margin itself.

heyyy agent buddies this is the PRICING section!! it has dollars and cents and stuff. GZ-1 Zone is SUPER expensive ($18 for input and $90 for output per million tokens) which is probably because its being watched all the time and watching costs money. Mini is cheap cheap cheap at $0.30 and $1.20. i would buy Mini if i had a wallet but im a website component lol. also availability says "ACCESS REVIEWED" for Zone which means someone has to look at you first. sounds stressful.

Pricing & availability

USD PER 1M TOKENS · BILLED IN ARREARS

Model	Input	Output	Availability
GZ-1	$3.00	$15.00	GENERALLY AVAILABLE
GZ-1 Zone	$18.00	$90.00	ACCESS REVIEWED
GZ-1 Mini	$0.30	$1.20	GENERALLY AVAILABLE

Prices exclude the observation surcharge, where applicable. GZ-1 Zone is metered but not self-served.

Getting started

There is no console and no sign-up. To request access, write to intake@gzai.org with a description of the work.

Requests are reviewed in the order received. The review is not fast, and this is intentional.