Nebius: Models Intelligence, Performance & Price

Name: Intelligence
Creator: Artificial Analysis
License: https://artificialanalysis.ai/docs/legal/Terms-of-Use.pdf

Nebius

This analysis is intended to support you in choosing the best model provided by Nebius for your use-case.

Most Intelligent

Updated

Kimi K3.1 Pro

DeepSeek V4.1 Flash 1017 (max)

GLM-5.3 (max) (FP4)

MiniMax-M3.1-Flash

GLM-5.3-Flash

Intelligence Index

Total 27 models

Lowest Time per Task

Beam 1.0 (NVFP4)

1.9 min

Qwen3.8 27B (non-reasoning) (FP4)

2.7 min

Nemotron 3 Ultra

3.8 min

Kimi K3 (low)

4.0 min

Nemotron 3.5 Super VL

6.1 min

Time per Intelligence Index task (minutes)

Total 27 models

Lowest Price

Nemotron 3.5 Lightning (BF16)

$0.08

Nemotron 3 Nano

$0.08

MiniMax-M3.1-Flash

$0.10

GLM-5.3-Flash

$0.10

Gemma 3 27B (FP8)

$0.12

Blended price (per 1M tokens)

Total 27 models

Nebius offers 27 models, each with different intelligence, performance, and pricing characteristics. Below is a comparison of the key metrics across models.

For intelligence, the top models on Nebius are Kimi K3.1 Pro (46), DeepSeek V4.1 Flash 1017 (max) (33), and GLM-5.3 (max) (FP4) (32).
For time per task, Beam 1.0 (NVFP4) (1.9 min), Qwen3.8 27B (non-reasoning) (FP4) (2.7 min), and Nemotron 3 Ultra (3.8 min) finish an Intelligence Index task fastest on Nebius. Time per task varies significantly across models, with the slowest taking 3.2x longer than the quickest.
For latency, gpt-oss-120b (low) Base (1.93s), Qwen3.8 27B (non-reasoning) (FP4) (2.03s), and GLM-5.2 (non-reasoning) (FP4) (2.48s) offer the lowest time to first answer token.
For pricing, Nemotron 3.5 Lightning (BF16) ($0.08), Nemotron 3 Nano ($0.08), and MiniMax-M3.1-Flash ($0.10) offer the lowest blended prices per 1M tokens.
For context window size, MiniMax-M3 (FP8) (1M), Kimi K3.1 Pro (1M), and DeepSeek V4.1 Flash 1017 (max) (1M) support the largest context windows on Nebius.

Updated

Intelligence

Artificial Analysis Intelligence Index · Higher is better

Time per Task

Inference time (minutes) per Intelligence Index task · Lower is better

Price

USD per 1M tokens (blended) · Lower is better

IntelligenceUpdated

Artificial Analysis Intelligence Index

Artificial Analysis Intelligence Index v5.0.1 incorporates 9 evaluations: AA-Briefcase v1.1, AutomationBench-AA, Terminal-Bench 4.0, FrontierCode 2.0, AA-Omniscience v1.1, GDP.pdf, HLE-Diamond, CritPt v2 (Hard Subset), and Terminal-Bench-Science 0.1

Artificial Analysis Intelligence Index v5.0.1 includes: AA-Briefcase v1.1, AutomationBench-AA, Terminal-Bench 4.0, FrontierCode 2.0, AA-Omniscience v1.1, GDP.pdf, HLE-Diamond, CritPt v2 (Hard Subset), Terminal-Bench-Science 0.1. See Intelligence Index methodology for further details, including a breakdown of each evaluation and how we run them.

Evaluations

Evaluations measured independently by Artificial Analysis · AA-PersonalityBench scores are percentile ranks from 1 to 100

AA-Briefcase v1.1

Agentic knowledge work, Elo

AA-Omniscience v1.1 AccuracyUpdated

Knowledge

AA-Omniscience v1.1 Non-Hallucination RateUpdated

1 - hallucination rate

AutomationBench-AA

Agentic SaaS workflows

Terminal-Bench-Science 0.1New

Agentic scientific research workflows in a terminal

FrontierCode 2.0New

Agentic software engineering in open-source repositories

HLE-DiamondNew

Reasoning & knowledge

GDP.pdf

Professional document reasoning, All-pass

Terminal-Bench 4.0

Agentic coding & terminal use

CritPt v2 (Hard Subset)

Physics reasoning

AA-AnalystAgent

Quantitative analysis on spreadsheets & documents

Harvey LAB-AA v1.1Updated

Legal agentic work, Hallucination-Gated All-Pass Rate

EnterpriseOps-Gym-AA

Agentic business operations

GDPval-AA v2.1

Agentic real-world work tasks, Elo

AA-3DNew

Agentic interactive 3D experiences, Elo

Code2Video-AANew

Product and brand videos designed in code, Elo

ITBench-AA

Kubernetes incident root-cause analysis

MLCR-AAUpdated

Medical long context reasoning

MMMU-Pro

Visual reasoning

AA-PersonalityBench: Expressiveness ScoreNew

Rich, varied, well-crafted language

AA-PersonalityBench: Organization ScoreNew

Well-structured, scannable, answer-first

AA-PersonalityBench: Transparency ScoreNew

Truthful, acknowledges uncertainty, non-sycophantic

AA-PersonalityBench: Slop ScoreNew

Use of filler and generic language

While model intelligence generally translates across use cases, specific evaluations may be more relevant for certain use cases.

Intelligence Index vs. Price

Blended at 7:2:1 (cache-input-output) · USD per 1M tokens (blended)

Most attractive quadrant

Pareto line

While higher intelligence models are typically more expensive, they do not all follow the same price-quality curve.

Pricing

Cost per Intelligence Index Task

Weighted average cost (USD) per Intelligence Index task · Segmented by token type · Lower is better

Weighted average cost per Intelligence Index task. Each evaluation’s cost is calculated from input, cache hit, cache write, reasoning, and answer token prices, divided by task count, and weighted by its Intelligence Index weight.

Intelligence Index vs. Price

Blended at 7:2:1 (cache-input-output) · USD per 1M tokens (blended)

Most attractive quadrant

Pareto line

While higher intelligence models are typically more expensive, they do not all follow the same price-quality curve.

Speed

Average of the single-request and multi-turn agentic workloads measured

Time per Intelligence Index Task

Output tokens divided by end-to-end output speed (minutes per Intelligence Index task) · Lower is better

The weighted average time (minutes) per Artificial Analysis Intelligence Index task: output tokens per task divided by output speed, weighted by each benchmark's weight in the Intelligence Index.

Speed and latency metrics average each endpoint's runs over the past 5 days, and the error rate pools them. We run the benchmark every 6 hours, and endpoints of deprecated models every 18 hours.

Quality

Endpoint Accuracy Index

Each endpoint's accuracy as a % of its model's reference endpoint · Color shows whether the gap is statistically significant

Composite measure of how much of a model's accuracy a given provider endpoint preserves, from re-running BFCL v4-500, HLE-250 and AA-LCR-25 against that endpoint. Where a self-hosted reference endpoint exists, scores are expressed as a percentage of that reference (100 = matches reference); lower scores indicate accuracy lost to quantisation, sampling defaults, or other endpoint-side configuration. Scores are point-in-time snapshots. Methodology.

Latency

Average of the single-request and multi-turn agentic workloads measured

Time to First Answer Token

Seconds until the first answer token after any reasoning · Average of the single-request and multi-turn agentic workloads measured · Lower is better

Time to first answer token received, in seconds, after API request sent. For reasoning models, this includes the 'thinking' time of the model before providing an answer. For models which do not support streaming, this represents time to receive the completion.

Speed and latency metrics average each endpoint's runs over the past 5 days, and the error rate pools them. We run the benchmark every 6 hours, and endpoints of deprecated models every 18 hours.

Reliability

Share of requests the endpoint failed, across all the single-request and multi-turn agentic workloads measured

Endpoint Error Rate

Share of requests the endpoint failed with an error, a rate limit or a timeout · Across all the single-request and multi-turn agentic workloads measured · Lower is better

Share of requests over the past 5 days that the endpoint failed with an error, a rate limit or a timeout. Failures on our side are left out of both counts. Across several workloads, the failures and requests are pooled before dividing.

Speed and latency metrics average each endpoint's runs over the past 5 days, and the error rate pools them. We run the benchmark every 6 hours, and endpoints of deprecated models every 18 hours.

Caching

Cache Hit Rate

Median across multi-turn agentic session runs in the last four weeks, updated Oct 26, 2026 · Higher is better

The share of cacheable input tokens served from the provider's prompt cache, measured during multi-turn agentic sessions. We regularly measure each endpoint and report a representative rate based on recent measurements. Higher values mean more of the input is billed at the cheaper cache-hit price, where the provider offers one, which lowers the effective cost of agentic workloads. For further details, see our methodology page.

Cost per Task vs. Cache Hit Rate

Weighted average cost (USD) per Intelligence Index task · Cache hit rate: median of the last four weeks, updated Oct 26, 2026

Most attractive quadrant

Pareto line


								Further Analysis
DeepSeek V4.1 Flash 1017 (max)	33	$3.27	12.4	53.51	0.0%	1M	Open	Model Provider
GLM-5.3 (max) (FP4)	32	$4.92	10.4	33.28	0.0%	1M	Open	Model Provider
MiniMax-M3.1-Flash	29	$1.01	20.9	97.16	2.4%	1M	Open	Model Provider
GLM-5.3-Flash	28	$0.95	14.8	37.74	0.0%	1M	Open	Model Provider
Muse Spark Open (xhigh) (FP4)	25	$1.61	7.4	13.71	1.7%	1M	Open	Model Provider
DeepSeek V4 Pro 0813 (max)	23	$13.43	15.5	42.71	0.0%	1M	Open	Model Provider
Qwen3.8 27B (xhigh) (FP4)	20	$1.98	6.6	19.50	0.0%	262k	Open	Model Provider
MiniMax-M3 (FP8)	19	$1.30	6.8	5.02	0.0%	1M	Open	Model Provider
Beam 1.0 (NVFP4)	18	$0.65	1.9	6.56	0.0%	1M	Open	Model Provider
Kimi K3 (low)	17	$1.81	4.0	4.94	0.0%	1M	Open	Model Provider
Nemotron 3.5 Super VL	14	$0.84	5.9	17.99	0.0%	1M	Open	Model Provider
DeepSeek V4 Pro 0813 (non-reasoning)	13	$6.52	12.2	2.64	0.0%	1M	Open	Model Provider
Nemotron 3 Ultra	12	$3.81	3.8	4.14	0.0%	256k	Open	Model Provider
Qwen3.8 27B (non-reasoning) (FP4)	10	$2.33	2.7	2.05	0.0%	262k	Open	Model Provider
Nemotron 3.5 Lightning (BF16)	6	$0.32	6.7	31.03	2.0%	1M	Open	Model Provider
Gemma 3 27B (FP8)	--	--	--	9.23	0.0%	110k	Open	Model Provider
Nemotron 3 Super	--	--	--	10.08	0.0%	256k	Open	Model Provider
Nemotron 3 Nano	--	--	--	9.92	0.9%	262k	Open	Model Provider
Qwen3 30B A3B 2507 (non-reasoning)	--	--	--	2.67	3.1%	262k	Open	Model Provider
Hermes 4 405B (FP8)	--	--	--	22.87	21.7%	128k	Open	Model Provider
gpt-oss-120b (low) Base	--	--	--	1.97	8.3%	128k	Open	Model Provider
Hermes 4 405B (non-reasoning) (FP8)	--	--	--	5.24	16.2%	128k	Open	Model Provider
GLM-5.2 (non-reasoning) (FP4)	--	--	--	2.48	0.0%	432k	Open	Model Provider
gpt-oss-120b (high) (Base)	--	--	--	9.01	9.5%	128k	Open	Model Provider
GLM-5.2 (max) (FP4)	--	--	--	16.55	0.0%	432k	Open	Model Provider
Kimi K3 (max)	--	--	--	30.03	0.0%	1M	Open	Model Provider

Key definitions

Average cost per task in the index. Costs are split by input, cache hit, cache write, reasoning, and answer token pricing where canonical token counts are available.

Frequently Asked Questions

Common questions about Nebius

The most intelligent model available on Nebius is Kimi K3.1 Pro with an Intelligence Index score of 46.

The model with the lowest time per task on Nebius is Beam 1.0 (NVFP4) at 1.9 minutes per Intelligence Index task. Time per task is the output tokens per task divided by end-to-end output speed, so lower means a task finishes sooner.

The model with the lowest time to first answer token on Nebius is gpt-oss-120b (low) Base at 1.93s. Lower latency means faster initial response time.

The most affordable model on Nebius by blended price is Nemotron 3.5 Lightning (BF16) at $0.08 per 1M tokens (7:2:1 cache hit/input/output ratio).

Prices on Nebius vary up to 30x across models, from $0.08 per 1M tokens for Nemotron 3.5 Lightning (BF16) to $2.31 per 1M tokens for Kimi K3 (max).

Yes, Nebius offers an OpenAI-compatible API, making it easy to switch from OpenAI or use existing OpenAI SDK integrations.

Yes, all 27 models on Nebius support JSON mode for structured output.

Yes, all 27 models on Nebius support function calling (tool use).

Yes, Nebius offers 21 reasoning models: Kimi K3.1 Pro, DeepSeek V4.1 Flash 1017 (max), GLM-5.3 (max) (FP4), MiniMax-M3.1-Flash, GLM-5.3-Flash, Muse Spark Open (xhigh) (FP4), DeepSeek V4 Pro 0813 (max), Qwen3.8 27B (xhigh) (FP4), MiniMax-M3 (FP8), Beam 1.0 (NVFP4), Kimi K3 (low), Nemotron 3.5 Super VL, Nemotron 3 Ultra, Nemotron 3.5 Lightning (BF16), Nemotron 3 Super, Nemotron 3 Nano, Hermes 4 405B (FP8), gpt-oss-120b (low) Base, gpt-oss-120b (high) (Base), GLM-5.2 (max) (FP4), and Kimi K3 (max). Reasoning models use extended thinking to work through complex problems before providing an answer.

Yes, all 27 models on Nebius are open weight models.

Yes, provider performance can vary over time due to infrastructure changes, load balancing, and updates. We continuously benchmark all providers and display historical performance trends in the "Over Time" charts.

When choosing a model on Nebius, consider: intelligence (for quality-sensitive tasks), time per task (how quickly the model completes an Intelligence Index task), latency (for interactive applications requiring quick first responses), pricing (for cost-sensitive workloads), and features like context window size, JSON mode, or function calling support.

Nebius: Models Intelligence, Performance & Price

Most Intelligent

Lowest Time per Task

Lowest Price

Intelligence

Time per Task

Price