Model Garden: how we measure AI models before they work for you

By Baybora Gülec · 5 min read

Dark blue background with points of light, the azena mark with “Model Garden”, a cross and the Nebius logo

TL;DR. The Model Garden is our own access point to AI models: one key for open models on Nebius Token Factory, for Claude and for GPT, with costs kept separate per customer. Which model takes on which task is not decided by leaderboards but by blind comparisons on real briefs. In our latest comparison, the open model Kimi K3 on Nebius built better presentations than GPT-6 Sol. Only Claude Opus 5.5 was slightly better.

One key for every model

The Model Garden has been live at api.azena.ai since October 7. With one key, our applications reach dozens of models: open models such as Kimi K3, DeepSeek or Qwen through Nebius Token Factory, plus Claude and GPT. Every request belongs to a customer, every customer has a monthly cap, and every response carries its cost in euros. In our test, the detour through the Garden added 1 to 7 milliseconds. Brand Studio has also been building its presentations through it since October 10.

Five names. The right model for every task.

You write the name of the task, not of the model. Which model does the work behind it is decided by azena.

Frontierazena/best

The best there is.

For hard tasks where every detail matters.

Today: Claude Fable 5.1

Frontierazena/code

For code and agents.

Programming, tools, long workflows.

Today: Claude Opus 5.5

Frontierazena/fast

Answers right away.

Chats, sorting, short texts.

Today: Claude Haiku 4.5

Open Sourceazena/open

Open and strong.

Analysis, drafts, long texts, at a fraction of the cost.

Today: DeepSeek V4 Pro

Open Sourceazena/cheap

For volume.

Sorting, extracting, summarising, thousands of times a day.

Today: gpt-oss-120b

When a better model arrives, we swap it in behind the name. Not a single line of your code changes.

This is what the Model Garden looks like.

Every plant is a model: the further left, the cheaper; the taller, the more text fits in.

Model Garden overview: models drawn as plants, arranged by price and context length
Table of redirects: DeepSeek V3.2 to DeepSeek V4 Pro, Kimi K2.5 to Kimi K3

Model changes without downtime

When a model is retired, the Garden routes old names to its successor. Your code keeps running.

List of API keys with status, monthly limit and requests per minute

Limits per key

Every key has its own monthly limit and its own rule for model changes.

Documentation quick start with a curl call to api.azena.ai

One address, done

Existing code in the OpenAI format only needs a new address and an azena key.

Screens from the Model Garden, with sample data.

How we measure

Leaderboards say little about how a model performs on your task. That is why we measure on real briefs. For presentations, this means: every model builds the same five decks through Brand Studio, with real slides, images and brand rules. A third model then puts two decks side by side and decides blind which one is better, in both orders.

Why not scores? Because scores are noisy. When we pitted the same model against itself, it scored 6.74 once and 6.34 the next time. A gap of 0.4 points therefore means nothing. The direct comparison separates cleanly.

What came out

Duels against Kimi K3 on Nebius · 5 briefs · 10 verdicts each

Verdict for the challengerVerdict for Kimi K3

Claude Opus 5.5 vs Kimi K36 : 4
GPT-5.6 Sol vs Kimi K30 : 10
GPT-6 Sol vs Kimi K30 : 10
Our own measurement from October 10, 2026: five real presentation briefs, every pair judged in both orders, each time by a third model.

Kimi K3 on Nebius won all ten verdicts against GPT-6 Sol and against GPT-5.6 Sol. Only Claude Opus 5.5 beat Kimi K3, at 6 to 4. In a first round with more models, DeepSeek V4 Pro won 35 percent and Claude Sonnet 5.5 20 percent of the verdicts against Kimi K3.

Price alone is misleading

What a deck costs and how long it takes

Cost per deck

GPT-6 Sol€0.40
GPT-5.6 Sol€0.76
Kimi K3€1.19
Claude Opus 5.5€1.42

Build time per deck

GPT-6 Sol35 s
GPT-5.6 Sol82 s
Kimi K395 s
Claude Opus 5.579 s
Same five briefs, average per deck. Scale: cost up to €1.50, time up to 100 seconds.

GPT-6 Sol builds a deck for 40 cents in 35 seconds. That is cheap and fast, but it loses every verdict. Kimi K3 costs €1.19 and takes 95 seconds, Claude Opus 5.5 costs €1.42 and takes 79 seconds. So the right question is not which model is cheapest, but what a usable result costs.

Why open source belongs in the mix.

Open models do not replace the frontier. They complement it wherever volume, cost and control matter.

33×

cheaper on output

gpt-oss-120b vs Claude Opus 5.5, list price per million tokens, October 2026.

10 : 0

against GPT-6 Sol

Kimi K3 on Nebius building presentations, judged blind.

Many providers

instead of one

Open weights run with many providers. If one gets more expensive, we switch and your code stays.

Your location

when it matters

If needed, an open model runs on dedicated servers or in your own data centre.

6 : 4

Frontier for the top end.

Where the last step in quality counts, Claude Opus 5.5 beats Kimi K3. So we use both: open models for breadth, frontier models for the top end, joined under one key in the Model Garden.

What you should do now

  1. Collect five real cases. Tasks from your daily work, not puzzles from the internet. Five are enough to start.
  2. Compare blind and in pairs. Two results side by side, without model names, in both orders. Scores alone deceive.
  3. Record cost and time per result. Only then do you see what a good answer really costs.
  4. Include open models. They keep up on many tasks and run with providers of your choice.

How we did this at scale is in our report on 353 runs with 13 models. If you want to know which model carries your task, we are happy to measure it with you.

A conversation worth having, even without an engagement.

Book an intro call

20 minutes, no obligation. Usually a personal reply on the same business day.

Related posts