Model benchmark

The leading models side by side. Measured independently.

From frontier models to open models for your own data center: how capable, how fast and how expensive the leading models are. Artificial Analysis measured them; we put them in context.

Talk to us Deployment and model choice

As of 26 October 2026Source: Artificial Analysis

Claude Opus 5.557.6Top model in the Intelligence Index

MiMo-V2.6-Pro46.3Strongest open model on the market, runs on-prem

Gemini 3.5 Flash335Tokens per second, the fastest model in this overview

Compare models

Strength, speed and price. Rarely all in one model.

Filter and sort by what matters for your use case: strength, speed, price or running in your own data center. Artificial Analysis measured all values independently, so you choose the model by your task, not by the vendor.

Sort by

47 of 47 models · sorted by Intelligence

Model comparison, source Artificial Analysis (artificialanalysis.ai), as of 26 October 2026
Rank	Model		Coding
01	Claude Opus 5.5Anthropic	Intelligence 57.6	Codingnot measured	Speed94Tokens/s	Price
02	Claude Sonnet 5.5AnthropicNew	Intelligence 56.0	Codingnot measured	Speed140Tokens/s	Price
03	Muse Wonder FullMetaNew	Intelligence 55.8	Codingnot measured	Speednot measured	Pricenot measured
04	Kimi K3.1 ProKimiNew	Intelligence 55.1	Codingnot measured	Speed48Tokens/s	Price
05	Claude Fable 5.1Anthropic	Intelligence 53.4	Coding81.6	Speed68Tokens/s	Price
06	GPT-6 AstraOpenAI	Intelligence 52.7	Coding76.9	Speed63Tokens/s	Price
07	Gemini 4 ArgonGoogleNew	Intelligence 52.6	Codingnot measured	Speed77Tokens/s	Price
08	Muse WonderMetaNew	Intelligence 52.4	Codingnot measured	Speed117Tokens/s	Price
09	GPT-6.1 SolOpenAINew	Intelligence 51.8	Codingnot measured	Speed58Tokens/s	Price
10	Claude Opus 5Anthropic	Intelligence 50.8	Coding78.0	Speednot measured	Price
11	GPT-5.6 SolOpenAI	Intelligence 47.0	Coding77.4	Speednot measured	Price
12	MiMo-V2.6-ProXiaomiOpen weightsOn-prem available	Intelligence 46.3	Codingnot measured	Speed42Tokens/s	Price
13	DeepSeek V4.1 Flash 1017DeepSeekNewOpen weightsOn-prem available	Intelligence 45.1	Codingnot measured	Speed152Tokens/s	Price
14	GLM-5.3Z AIOpen weightsOn-prem available	Intelligence 44.8	Coding74.8	Speed86Tokens/s	Price
15	Kimi K3KimiOpen weightsOn-prem available	Intelligence 43.6	Coding76.2	Speed37Tokens/s	Price
16	GPT-5.6 TerraOpenAI	Intelligence 42.1	Coding76.7	Speed121Tokens/s	Price
17	Claude Opus 4.8Anthropic	Intelligence 41.8	Coding74.3	Speednot measured	Price
18	GLM-5.3-FlashZ AIOpen weightsOn-prem available	Intelligence 41.8	Coding71.5	Speed65Tokens/s	Price
19	Gemini 3.8 FlashGoogle	Intelligence 40.9	Coding76.3	Speed120Tokens/s	Price
20	Claude Opus 4.7Anthropic	Intelligence 40.7	Coding73.6	Speednot measured	Price
21	GPT-5.4OpenAI	Intelligence 39.0	Coding71.1	Speednot measured	Price
22	GPT-5.5OpenAI	Intelligence 38.4	Coding74.9	Speed151Tokens/s	Price
23	Claude Sonnet 5Anthropic	Intelligence 38.2	Coding71.5	Speednot measured	Price
24	GPT-5.6 LunaOpenAI	Intelligence 37.3	Coding71.4	Speednot measured	Price
25	Qwen3.8 27BAlibabaOpen weightsOn-prem available	Intelligence 33.7	Coding68.1	Speed47Tokens/s	Price
26	Gemini 3.5 FlashGoogle	Intelligence 33.6	Codingnot measured	Speed378Tokens/s	Price
27	Claude Opus 4.6Anthropic	Intelligence 31.9	Codingnot measured	Speednot measured	Price
28	Claude Sonnet 4.6Anthropic	Intelligence 30.1	Coding63.0	Speednot measured	Price
29	Gemini 3 ProGoogle	Intelligence 28.0	Codingnot measured	Speednot measured	Price
30	Nemotron 3.5 Super VLNVIDIANewOpen weightsOn-prem available	Intelligence 26.5	Codingnot measured	Speed456Tokens/s	Price
31	GPT-5.1OpenAI	Intelligence 24.7	Coding49.4	Speednot measured	Price
32	GPT-5OpenAI	Intelligence 23.0	Coding37.8	Speednot measured	Price
33	Nemotron 3 Ultra 550B A55BNVIDIAOpen weightsOn-prem available	Intelligence 22.9	Coding49.3	Speed172Tokens/s	Price
34	Qwen3.6 27BAlibabaOpen weightsOn-prem available	Intelligence 21.4	Coding53.7	Speednot measured	Price
35	Gemini 2.5 ProGoogle	Intelligence 16.1	Coding33.3	Speednot measured	Price
36	Gemma 4 31BGoogleOpen weightsOn-prem available	Intelligence 14.7	Coding43.4	Speed63Tokens/s	Pricenot measured
37	Qwen3 VL 235BAlibabaOpen weightsOn-prem available	Intelligence 13.4	Codingnot measured	Speednot measured	Pricenot measured
38	Gemini 2.5 FlashGoogle	Intelligence 13.1	Codingnot measured	Speednot measured	Price
39	GPT-4.1OpenAI	Intelligence 12.7	Codingnot measured	Speednot measured	Price
40	gpt-oss-120bOpenAIOpen weightsOn-prem available	Intelligence 11.6	Coding30.4	Speed262Tokens/s	Price
41	gpt-oss-20bOpenAIOpen weightsOn-prem available	Intelligence 10.0	Codingnot measured	Speed208Tokens/s	Price
42	Mistral Large 3MistralOpen weightsOn-prem available	Intelligence 9.3	Coding20.1	Speednot measured	Price
43	Gemini 2.5 Flash-LiteGoogle	Intelligence 8.5	Codingnot measured	Speednot measured	Price
44	GPT-4oOpenAI	Intelligence 8.4	Codingnot measured	Speednot measured	Price
45	Llama 3.3 70BMetaOpen weightsOn-prem available	Intelligence 7.7	Coding11.9	Speed96Tokens/s	Price
46	GPT-4o miniOpenAI	Intelligence 6.7	Coding11.4	Speednot measured	Price
47	Gemma 3 27BGoogleOpen weightsOn-prem available	Intelligence 4.9	Coding10.1	Speednot measured	Pricenot measured

Source: Artificial Analysis (artificialanalysis.ai), as of 26 October 2026. For models with adjustable reasoning effort, the measurement at high or maximum reasoning effort applies. If a value is missing, Artificial Analysis did not measure it; we do not estimate.

Performance vs. price

Each dot is a model: the higher, the stronger; the further right, the more expensive. The efficiency frontier connects the models that deliver more than any cheaper one.

Anthropic
Meta
Kimi
OpenAI
Google
Xiaomi
DeepSeek
Z AI
Alibaba
NVIDIA
Mistral

Intelligence Index

Proprietary
Open weights
Efficiency frontier

Hover over a dot or click it: name and values appear in the panel below.

Claude Opus 5.5

Anthropic · released Sep 22, 2026

Intelligence: 57.6
Coding: not measured
Speed: 94Tokens/s
Price

Not plotted (price or index not measured): Muse Wonder Full, Gemma 4 31B, Qwen3 VL 235B, Gemma 3 27B. Source: Artificial Analysis (artificialanalysis.ai), as of 26 October 2026.

How to read the values

Intelligence Index

The independent rating by Artificial Analysis combines several tests of knowledge, reasoning, math and coding into a single number. Higher is better. The Coding Index measures coding alone.

Speed

How many tokens a model generates per second, measured at the provider. A token is part of a word. In your own data center, speed depends on your hardware.

Price tier

The list price per million tokens at the providers, grouped into five tiers. More bars means more expensive. Your terms depend on the deployment and are stated in your quote.

On-prem capable

The model weights are openly published. The model can run on your own infrastructure or in a sovereign environment.

Open weights and on-prem

Models with open weights.

Can run in your own data center.

Models with open weights.

Can run in your own data center.

Many of these models are published with open weights. For organizations such as banks, insurers, law firms, hospitals or public authorities, this means the model runs where the data lives.

Performance

46.3 vs. 57.6

The strongest open model on the market, MiMo-V2.6-Pro, reaches about four fifths of the score of the strongest proprietary model, Claude Opus 5.5, in the Intelligence Index.

Weights

Openly published

The model weights are publicly available, for example on Hugging Face. You decide when a model changes, not the vendor.

Deployment

In-house or sovereign

In your data center, in a sovereign cloud such as Schwarz Digits Cloud (STACKIT) or hybrid, depending on your protection needs.

Requests

Never leave your network

When the neuland.ai HUB and the open model run in your data center, requests and responses never leave your network.

See the architecture

Model choice

You decide which model does the work.

Per group, per task, with your own key if you want.

You decide which model does the work.

Per group, per task, with your own key if you want.

You decide in the neuland.ai HUB which model handles a task. Admins approve models per group; Auto mode or the choice in chat handles the rest.

Auto mode

Each message goes to a suitable model based on its complexity, only among those approved for the group. In Sales, a fast model summarizes the call notes, while the strongest model the group may use analyzes a long annual report.

Free choice in chat

Anyone who wants to can choose for themselves, even mid-chat, and compare the answers of two models side by side.

Bring your own keys

With Bring Your Own Key (BYOK), you use existing contracts with model providers: you store your own API key in the neuland.ai HUB, and model usage then runs through your own contract.

Switch without rebuilding

When a stronger model arrives in the neuland.ai HUB, admins approve it for a group. From then on, the group’s agents work with it, without anyone rebuilding them.

Deployment and model choice

FAQ

What customers ask first about the models.

Which model is the best?

It depends on the task. Claude Opus 5.5 currently leads the Intelligence Index (57.6), closely followed by Claude Sonnet 5.5 (56.0). For a quick summary, a model like Gemini 3.5 Flash at 335 tokens per second, the fastest in this overview, is often enough, and gpt-oss-120b sits in the lowest price tier and even runs in your own data center. Auto mode makes this trade-off for every message.

Which model is the best?

Which models can run on-prem?

All models with open weights, marked as on-prem capable in the overview. These include Qwen3.8 27B, gpt-oss-120b, Mistral Large 3, Llama 3.3 70B and Gemma 4 31B, as well as stronger open models such as MiMo-V2.6-Pro, GLM-5.3, Kimi K3 or DeepSeek V4.1 Flash. Proprietary models such as Claude, GPT or Gemini are connected via their provider’s cloud. We work out with you which hardware an open model needs as part of the deployment concept.

Which models can run on-prem?

Is our data used for training?

No. Your data is never used to train language models and stays within your company context, regardless of which model you use.

Is our data used for training?

No. Your data is never used to train language models and stays within your company context, regardless of which model you use.

Can we use our own API keys?

Yes, via Bring Your Own Key (BYOK). If you already have a contract with a model provider, you store your own API key in the neuland.ai HUB. Model usage then runs through your own contract.

Can we use our own API keys?

Yes, via Bring Your Own Key (BYOK). If you already have a contract with a model provider, you store your own API key in the neuland.ai HUB. Model usage then runs through your own contract.

How current are the values?

As of 26 October 2026. The values come from Artificial Analysis, an independent provider of model comparisons that continuously updates its measurements. If a value is missing, Artificial Analysis has not measured it for this model; we do not estimate.

How current are the values?

Why don’t you list prices?

Because the price of a request in the neuland.ai HUB depends on the deployment: through a provider’s cloud, through your own contract or on your own hardware, the math is different each time. The price tier therefore only shows how expensive the models are relative to each other, based on the providers’ list prices per million tokens. We prepare your quote based on your use case; details in the pricing overview.

Why don’t you list prices?

What does the Intelligence Index measure?

The independent rating by Artificial Analysis combines several tests of knowledge, reasoning, math and coding into a single number. Higher is better. The Coding Index measures coding alone.

What does the Intelligence Index measure?

The independent rating by Artificial Analysis combines several tests of knowledge, reasoning, math and coding into a single number. Higher is better. The Coding Index measures coding alone.

What does speed tell me?

How many tokens a model generates per second, measured at the provider. A token is part of a word. In your own data center, speed depends on your hardware.

What does speed tell me?

How many tokens a model generates per second, measured at the provider. A token is part of a word. In your own data center, speed depends on your hardware.

Who decides which models my group may use?

Admins approve models per group. Anyone who wants to can then choose in chat, even mid-conversation, and compare the answers of two models side by side; otherwise Auto mode makes the choice.

Who decides which models my group may use?

Admins approve models per group. Anyone who wants to can then choose in chat, even mid-conversation, and compare the answers of two models side by side; otherwise Auto mode makes the choice.

What does open weights mean?

The model weights are publicly released, for example on Hugging Face. Such models can run on your own infrastructure or in a sovereign environment.

What does open weights mean?

The model weights are publicly released, for example on Hugging Face. Such models can run on your own infrastructure or in a sovereign environment.

The right model for each task.

We choose it with you.

You bring your tasks, we put the numbers next to them: which model is strong enough, which is fast enough and which has to run in your own data center.

Talk to us

After your request

You name your tasks and protection needs

We assign suitable models to each task

You receive the selection per department as the basis for your onboarding plan