What is the cheapest AI21 model?

There are no tracked model rates available to establish a cheapest option.

Which AI21 model has the highest AA Intelligence score?

No Artificial Analysis Intelligence Index results are currently listed for AI21. Missing results do not establish that a model is worse than a scored model.

How do I calculate monthly AI21 API costs?

Multiply input tokens in millions by the input rate and add output tokens in millions multiplied by the output rate. This excludes tool charges, cache-specific charges, taxes and any endpoint-specific adjustments.

Which AI21 model has the largest context window?

No model context limits are currently listed for this provider.

What is the latest AI21 model in the catalog?

No usable release dates are currently recorded, so this catalog cannot establish the latest model.

Do AI21 models support image inputs?

Image-input support is not currently listed for the tracked models. Missing capability data is not proof that image input is unavailable; confirm with the provider's documentation.

Can prompt caching lower AI21 API costs?

No cached-input rates are currently listed for this provider's tracked models. Check endpoint documentation for caching eligibility and charges rather than assuming a discount.

Do AI21 models support tool calling?

Tool-calling capabilities are not currently listed for these models. Unlisted support is not the same as confirmed lack of support; check each model and the endpoint documentation.

How should I compare AI21 benchmark scores and prices?

Compare scores within the same evaluation: Artificial Analysis Intelligence Index, DesignArena Elo, general Arena Elo and MMLU-Pro use different tasks and scales. Missing results are not zeroes. Pair a relevant measured benchmark with input/output pricing and tests of your own workload; the cheapest or highest-scoring model is not automatically the best fit.

Are retired AI21 models included in the price list?

The price list shows active models by default. Turn on Show retired models to include retired catalog entries, where available. A retained model page preserves reference information, not a promise that its API is still available.

Is AI21 an AI lab or an inference host?

AI21 is categorized as an inference host in this catalog. Hosts provide access to models; check each model page for its creator and listed endpoints.

What are AI21 Labs' flagship models?

AI21 Labs' flagship models include the Jurassic-2 family and the Jamba family. The Jamba models, particularly Jamba 1.5 Large and Mini, leverage a hybrid Mamba SSM and Transformer architecture for efficiency and long context processing.

What is the maximum context window supported by AI21 Labs models?

The Jamba family of models from AI21 Labs supports an industry-leading context window of up to 256,000 tokens. This allows for processing extensive documents and complex data.

How is AI21 Labs' pricing structured?

Pricing for AI21 Labs' models is generally calculated per token, with separate rates for input and output tokens. Costs can vary based on whether models are accessed directly via AI21 Studio or through third-party cloud hosts like AWS Bedrock or Google Cloud.

Can I deploy AI21 Labs models on my own infrastructure?

Yes, AI21 Labs offers open-weight models from the Jamba family for private deployment, including VPC and on-premises environments. These models are also available on platforms like Hugging Face, Kaggle, GCP Model Garden (self-deploy), and Microsoft Foundry (self-deploy).

What is AI21 Maestro?

AI21 Maestro is an advanced AI system for rapidly creating and deploying knowledge agents that automate data-intensive business tasks. It includes Retrieval-Augmented Generation (RAG) capabilities and is optimized for searching, reasoning, validating, and adapting in real-time.

Does AI21 Labs offer specialized models for specific tasks?

Yes, AI21 Labs provides Task-Specific Models (TSMs) such as Contextual Answers, Summarize, Paraphrase, and Grammatical Error Correction. These models are built to deliver out-of-the-box value and higher accuracy for common commercial use cases.

Inference host

AI21 API pricing & models

www.ai21.com

AI21 Labs is an Israeli company specializing in natural language processing (NLP), developing advanced AI systems and foundation models designed for enterprise applications. Their mission is to empower businesses with state-of-the-art Large Language Models (LLMs) and AI applications to redefine how humans interact with text. AI21 serves open-weight models from other labs through its own API.

Overview

About AI21

The company offers a suite of models and tools, including the Jurassic-2 family and the Jamba family. The Jamba models are particularly notable for their hybrid Mamba State Space Model (SSM) and Transformer architecture, which enables efficient processing of exceptionally long context windows.

AI21 Labs provides access to its models through its AI21 Studio developer platform, which includes a Python SDK and REST API, and also via major cloud providers like Amazon Bedrock and Google Cloud Vertex AI. They focus on solutions for high-value, data-intensive workflows such as grounded question answering, RAG workflows, and agentic systems, aiming to reduce hallucinations and improve accuracy for business-critical tasks.

Long-context Retrieval-Augmented GenerationGrounded question answeringFinancial document analysisAutomated content summarizationEnterprise knowledge agentsMultilingual text generation

Headquarters: ILPrivacy policy Terms Status page

Strengths

Jamba models offer an industry-leading context window of up to 256,000 tokens, enabling processing of massive documents for enterprises.
The hybrid Mamba SSM and Transformer architecture of Jamba models provides high inference speed and efficiency, especially for long contexts.
AI21 Labs provides specialized Task-Specific Models (TSMs) for common business needs like summarization, paraphrasing, and grammatical error correction.
Maestro, an advanced AI system, helps deploy knowledge agents that automate data-intensive tasks, featuring RAG capabilities and self-validation to reduce hallu
Models are available across multiple cloud platforms (AWS Bedrock, Google Cloud Vertex AI) and open-source hubs (Hugging Face, Kaggle) for flexible deployment.
The platform supports advanced developer features like function calling and structured JSON output for integrating AI into complex applications.

Limitations

Older model families, such as Jurassic-2 Ultra, have a significantly smaller context window (8,191 tokens) compared to the newer Jamba models.
While Jamba's hybrid architecture is efficient, pure Transformer implementations can sometimes be faster for very short sequence lengths (below 2,000-4,000 toke
Access to some models or higher rate limits may require contacting sales, indicating potential non-standardized access for specific enterprise needs.

Open-weight models

Models you can run on AI21

AI21 hosts popular open-weight models. Dedicated AI21 pricing is coming soon — here are the models it typically serves.

Show retired models


Muse Spark 1.3 Meta	$1.25	$4.25	1M	48.1	1358	—	—	0%
Muse Spark 1.2 Meta	$1.25	$4.25	1M	39.6	1318	—	—	0%
DeepSeek V4.1 Flash DeepSeek	$0.05	$1.2	1M	39.5	1325	—	—	357%
DeepSeek V4 Pro 0423 DeepSeek	$0.29232	$0.58464	1M	36	1244	—	—	-19%
DeepSeek V4 Flash Vision Exp DeepSeek	$0.2156	$0.6468	1M	34.8	—	—	—	96%
DeepSeek V4 Flash 0423 DeepSeek	$0.03	$0.056	1M	34.3	1209	—	—	3%
Muse Spark 1.1 Meta	$1.25	$4.25	1M	33.7	—	—	—	0%
DeepSeek V3.2 DeepSeek	$0.259	$0.42	164K	16	—	—	—	48%
DeepSeek V3.2 Exp DeepSeek	$0.27	$0.41	164K	16	—	—	—	0%
DeepSeek V3.1 Terminus DeepSeek	$0.27	$1	164K	13.9	—	—	—	0%
DeepSeek V3.1 DeepSeek	$0.25	$0.95	164K	13.7	—	—	—	0%
DeepSeek R1 DeepSeek	$0.7	$2.5	64K	13.1	—	1239	80	3%
Llama 4 Maverick Meta	$0.1875	$0.6525	1M	10	883	—	—	0%
DeepSeek V3 0324 DeepSeek	$0.29	$1.14	164K	9.7	—	—	—	167%
DeepSeek V3 DeepSeek	$0.2574	$1.0287	164K	8.5	—	1211	75.4	-20%
Llama 4 Scout Meta	$0.1	$0.3	1.3M	8.1	793	—	—	0%
R1 Distill Llama 70B DeepSeek	$0.8	$0.8	8K	7.9	—	—	—	0%
Llama 3.3 70B Instruct Meta	$0.1	$0.32	131K	7.7	—	—	—	107%
Llama 3.1 8B Instruct Meta	$0.05	$0.08	131K	6.9	—	—	—	0%
Llama 3.1 70B Instruct Meta	$0.4	$0.4	131K	6.6	—	—	—	0%
Llama 3.2 3B Instruct Meta	$0.05	$0.33	131K	5.7	—	—	—	0%
Llama 3.2 1B Instruct Meta	$0.027	$0.201	60K	4.8	—	—	—	0%
DeepSeek V4 Flash 0731 DeepSeek	$0.018	$0.32	1M	—	1234	—	—	323%
DeepSeek V4 Pro 0813 DeepSeek	$0.66	$1.98	1M	—	—	—	—	282%
Llama Guard 4 12B Meta	$0.18	$0.18	164K	—	—	—	—	0%
Muse Glimmer 30B Meta	$0.3	$1.2	131K	—	—	—	—	133%
Muse Spark 1.2 Contributor Meta	$0.1	$0.2	1M	—	—	—	—	0%
Muse Spark 1.3 Contributor Meta	$0.1	$0.2	1M	—	—	—	—	0%
R1 0528 DeepSeek	$0.5	$2.15	164K	—	—	—	—	0%

Muse Spark 1.3 Meta

Input /1M: $1.25
Output /1M: $4.25
Context: 1M
AA Intelligence: 48.1
DesignArena Elo: 1358
Arena Elo: —
MMLU-Pro: —
30d trend: 0%

Muse Spark 1.2 Meta

Input /1M: $1.25
Output /1M: $4.25
Context: 1M
AA Intelligence: 39.6
DesignArena Elo: 1318
Arena Elo: —
MMLU-Pro: —
30d trend: 0%

DeepSeek V4.1 Flash DeepSeek

Input /1M: $0.05
Output /1M: $1.2
Context: 1M
AA Intelligence: 39.5
DesignArena Elo: 1325
Arena Elo: —
MMLU-Pro: —
30d trend: 357%

DeepSeek V4 Pro 0423 DeepSeek

Input /1M: $0.29232
Output /1M: $0.58464
Context: 1M
AA Intelligence: 36
DesignArena Elo: 1244
Arena Elo: —
MMLU-Pro: —
30d trend: -19%

DeepSeek V4 Flash Vision Exp DeepSeek

Input /1M: $0.2156
Output /1M: $0.6468
Context: 1M
AA Intelligence: 34.8
DesignArena Elo: —
Arena Elo: —
MMLU-Pro: —
30d trend: 96%

DeepSeek V4 Flash 0423 DeepSeek

Input /1M: $0.03
Output /1M: $0.056
Context: 1M
AA Intelligence: 34.3
DesignArena Elo: 1209
Arena Elo: —
MMLU-Pro: —
30d trend: 3%

Muse Spark 1.1 Meta

Input /1M: $1.25
Output /1M: $4.25
Context: 1M
AA Intelligence: 33.7
DesignArena Elo: —
Arena Elo: —
MMLU-Pro: —
30d trend: 0%

DeepSeek V3.2 DeepSeek

Input /1M: $0.259
Output /1M: $0.42
Context: 164K
AA Intelligence: 16
DesignArena Elo: —
Arena Elo: —
MMLU-Pro: —
30d trend: 48%

DeepSeek V3.2 Exp DeepSeek

Input /1M: $0.27
Output /1M: $0.41
Context: 164K
AA Intelligence: 16
DesignArena Elo: —
Arena Elo: —
MMLU-Pro: —
30d trend: 0%

DeepSeek V3.1 Terminus DeepSeek

Input /1M: $0.27
Output /1M: $1
Context: 164K
AA Intelligence: 13.9
DesignArena Elo: —
Arena Elo: —
MMLU-Pro: —
30d trend: 0%

DeepSeek V3.1 DeepSeek

Input /1M: $0.25
Output /1M: $0.95
Context: 164K
AA Intelligence: 13.7
DesignArena Elo: —
Arena Elo: —
MMLU-Pro: —
30d trend: 0%

DeepSeek R1 DeepSeek

Input /1M: $0.7
Output /1M: $2.5
Context: 64K
AA Intelligence: 13.1
DesignArena Elo: —
Arena Elo: 1239
MMLU-Pro: 80
30d trend: 3%

Llama 4 Maverick Meta

Input /1M: $0.1875
Output /1M: $0.6525
Context: 1M
AA Intelligence: 10
DesignArena Elo: 883
Arena Elo: —
MMLU-Pro: —
30d trend: 0%

DeepSeek V3 0324 DeepSeek

Input /1M: $0.29
Output /1M: $1.14
Context: 164K
AA Intelligence: 9.7
DesignArena Elo: —
Arena Elo: —
MMLU-Pro: —
30d trend: 167%

DeepSeek V3 DeepSeek

Input /1M: $0.2574
Output /1M: $1.0287
Context: 164K
AA Intelligence: 8.5
DesignArena Elo: —
Arena Elo: 1211
MMLU-Pro: 75.4
30d trend: -20%

Showing 15 of 29

15 of 29 models

AI21 Labs' Jamba Models: Hybrid Architecture and Long Context

The Jamba family of models represents a significant offering from AI21 Labs, distinguishing itself with a unique hybrid Mamba State Space Model (SSM) and Transformer architecture. This innovative design combines the strengths of both paradigms: the linear scaling and efficiency of Mamba models for processing long sequences, and the strong performance and in-context learning capabilities of Transformers. As a result, Jamba models are capable of handling an extensive context window of up to 256,000 tokens, which is among the longest available for open models, making them highly suitable for applications requiring deep analysis of large documents, such as legal contracts, research papers, or financial reports.

Available in different sizes like Jamba 1.5 Mini and Jamba 1.5 Large, these models also incorporate a Mixture of Experts (MoE) component, which contributes to their inference efficiency. They are optimized for tasks like long-context Retrieval-Augmented Generation (RAG), grounded question answering, and data classification. Developers can access Jamba models through AI21 Studio, as well as managed services on Amazon Bedrock and Google Cloud Vertex AI, and can also deploy open-w

Enterprise Solutions and Task-Specific Capabilities

AI21 Labs focuses on providing robust AI solutions tailored for enterprise workflows. Beyond their foundational Jamba models, they offer AI21 Maestro, an advanced system designed to build and deploy intelligent agents that automate complex, data-intensive business tasks. Maestro features built-in RAG capabilities, semantic search, and web search integration, enabling agents to retrieve, reason, validate, and adapt information in real-time while adhering to performance and cost constraints. This system is specifically engineered to improve accuracy and significantly reduce AI hallucinations in business-critical applications.

For more specialized needs, AI21 Labs also provides a suite of Task-Specific Models (TSMs). These include Contextual Answers for accurate information retrieval from internal knowledge bases, Summarize for condensing lengthy texts, Paraphrase for rewriting content, and Grammatical Error Correction. These TSMs are designed to offer immediate value and higher accuracy for common commercial use cases, requiring less prompt engineering and allowing organizations to quickly integrate generative AI features into their products and services.

FAQ