Top 10 Cheapest LLM APIs in 2026

Looking for powerful LLMs that won’t drain your budget? We ranked the 10 cheapest LLM API models available on Novita AI in 2026, with pricing starting at just $0.02 per million tokens. From Meta’s Llama 3.1 8B to Alibaba’s Qwen3 Coder, these models cover general chat, reasoning, code generation, multilingual support, and long-context tasks — all at a fraction of what premium models cost. Our top three picks: Llama 3.1 8B Instruct ($0.02/M), Qwen3 4B ($0.03/M), and Llama 3 8B Instruct ($0.04/M).

How We Ranked These Models

We selected models based on three criteria:

1. Price — Input cost per million tokens on Novita AI, ranked from lowest to highest.
2. Practical utility — The model must be useful for real-world tasks (general chat, code generation, reasoning, or tool use), not just cheap.
3. Availability — All models are live on Novita AI’s Serverless endpoints and accessible via OpenAI-compatible API right now.

We excluded OCR-only models, dedicated endpoints, and highly specialized tools that don’t function as general-purpose LLMs.

The 10 Cheapest LLM APIs on Novita AI

1. Meta Llama 3.1 8B Instruct

Spec	Detail
Developer	Meta
Parameters	8B
Context Length	16K
Pricing (Input / Output)	$0.02 / $0.05 per M tokens
Quantization	FP8
Best For	General chat, content generation, lightweight tasks

Meta’s Llama 3.1 8B Instruct is the most affordable general-purpose LLM you can access via API today. Trained on over 15 trillion tokens and fine-tuned with supervised learning and RLHF, this 8B-parameter model punches well above its weight — outperforming several closed-source models on industry benchmarks despite its compact size.

At just $0.02 per million input tokens on Novita AI, it’s the go-to choice for developers who need a reliable, fast LLM for chat applications, content generation, and simple instruction-following tasks without spending more than pocket change.

Pros

• Lowest price on this list at $0.02/M input tokens on Novita AI.
• Strong general performance for an 8B model.
• Proven and battle-tested across thousands of production deployments.

Cons

• 16K context window is limited compared to newer models.
• Text-only — no multimodal capabilities.

Best For

Budget-conscious developers who need a dependable, general-purpose LLM for high-volume, low-complexity tasks.

2. Qwen3 4B

Spec	Detail
Developer	Alibaba (Qwen Team)
Parameters	4B
Context Length	128K
Pricing (Input / Output)	$0.03 / $0.03 per M tokens
Quantization	FP8
Best For	Long-document processing, creative writing, role-playing

Qwen3 4B delivers a remarkable combination on Novita AI: 128K context length at just $0.03 per million tokens for both input and output. That’s the longest context window in this price range by a wide margin.

Despite having only 4 billion parameters, it supports both reasoning and non-reasoning modes with seamless switching during conversations. The model shows strong performance in creative writing, role-playing, multi-turn dialogue, and instruction following — making it far more versatile than its size suggests.

Pros

• 128K context at $0.03/M on Novita AI — unmatched value for long-document tasks.
• Identical input and output pricing simplifies cost estimation.
• Supports tool calling and reasoning modes.

Cons

• 4B parameters limits performance on complex reasoning tasks.
• Max output capped at 20K tokens.

Best For

Developers who need to process long documents, conversation histories, or large code files on a tight budget.

3. Meta Llama 3 8B Instruct

Spec	Detail
Developer	Meta
Parameters	8B
Context Length	8K
Pricing (Input / Output)	$0.04 / $0.04 per M tokens
Quantization	BF16
Best For	Simple dialogue, content generation, balanced pricing

Llama 3 8B Instruct is the predecessor to 3.1 and remains a popular choice for its flat, predictable pricing — $0.04 per million tokens for both input and output on Novita AI. This makes cost estimation dead simple for high-volume workloads.

Optimized for dialogue use cases, it delivers strong performance compared to leading closed-source models in human evaluations. The 8K context window is shorter than newer models, but for straightforward chat, Q&A, and content generation tasks, it’s more than enough.

Pros

• Flat $0.04/M pricing for both input and output on Novita AI — simplest cost model.
• Strong dialogue performance validated by human evaluations.
• Mature, well-documented model with a massive ecosystem.

Cons

• 8K context window — the shortest on this list.
• No reasoning mode or tool calling support.

Best For

Teams who want predictable costs with flat input/output pricing for simple, high-volume chat and generation tasks.

4. OpenAI GPT-OSS 20B

Spec	Detail
Developer	OpenAI
Parameters	21B (3.6B active, MoE)
Context Length	131K
Pricing (Input / Output)	$0.04 / $0.15 per M tokens
Quantization	FP4
Best For	Reasoning, tool use, agentic workflows

GPT-OSS 20B is OpenAI’s entry into the open-weight arena — a 21B-parameter Mixture-of-Experts model released under the Apache 2.0 license. With only 3.6B active parameters per forward pass, it’s designed for low-latency inference while delivering reasoning capabilities that rival much larger models.

The model supports configurable reasoning depth, function calling, tool use, structured outputs, and JSON mode — making it one of the most feature-rich cheap models on this list. At $0.04/M input tokens on Novita AI, you’re getting OpenAI-grade reasoning for a fraction of what GPT-4o costs.

Pros

• OpenAI quality at open-source pricing on Novita AI.
• MoE architecture — only 3.6B active params for fast inference.
• Full support for tool use, function calling, and structured outputs.

Cons

• Relatively new — smaller community ecosystem compared to Llama.
• MoE models can have less consistent output quality on niche tasks.

Best For

Developers building agentic applications who want OpenAI-level reasoning at a fraction of the cost on Novita AI.

5. Mistral Nemo

Spec	Detail
Developer	Mistral AI × NVIDIA
Parameters	12B
Context Length	60K
Pricing (Input / Output)	$0.04 / $0.17 per M tokens
Quantization	FP8
Best For	Multilingual applications, function calling

Mistral Nemo is a 12B-parameter model built through a collaboration between Mistral AI and NVIDIA. It supports 11 languages — English, French, German, Spanish, Italian, Portuguese, Chinese, Japanese, Korean, Arabic, and Hindi — making it the strongest multilingual option in this price range on Novita AI.

With a 60K context window, function calling support, and structured output capabilities, it’s a well-rounded model that handles multilingual chat, translation, and document processing tasks with ease. At $0.04/M input tokens on Novita AI, it’s one of the most cost-effective ways to serve a global user base.

Pros

• 11-language support — the best multilingual model under $0.05/M on Novita AI.
• Built with NVIDIA — optimized for efficient inference.
• Function calling and structured output support.

Cons

• 60K context — shorter than Qwen3 or GPT-OSS models.
• No reasoning mode.

Best For

Teams building multilingual products on Novita AI that need reliable language support across diverse markets.

6. OpenAI GPT-OSS 120B

Spec	Detail
Developer	OpenAI
Parameters	117B (5.1B active, MoE)
Context Length	131K
Pricing (Input / Output)	$0.05 / $0.25 per M tokens
Quantization	FP4
Best For	High-reasoning tasks, production agentic systems

GPT-OSS 120B is the big sibling — a 117B-parameter MoE model that activates only 5.1B parameters per forward pass, designed to run on a single H100 GPU. It delivers production-grade reasoning, full chain-of-thought access, configurable reasoning depth, and native tool use including function calling and browsing.

At $0.05/M input tokens on Novita AI, this is arguably the most powerful LLM you can get for under a dime per million tokens. It’s the model to choose when your task demands serious reasoning capability but your budget says “no” to GPT-4o pricing.

Pros

• 117B parameters with only 5.1B active — massive capability, efficient inference.
• Full tool use on Novita AI: function calling, browsing, structured outputs.
• Configurable reasoning depth for cost/quality trade-offs.

Cons

• Output pricing ($0.25/M) is higher than simpler models on this list.
• MoE models may underperform dense models of similar total size on some tasks.

Best For

Production AI systems on Novita AI that need high reasoning power at scale without the cost of premium closed-source APIs.

7. Qwen 2.5 7B Instruct

Spec	Detail
Developer	Alibaba (Qwen Team)
Parameters	7B
Context Length	32K
Pricing (Input / Output)	$0.07 / $0.07 per M tokens
Quantization	BF16
Best For	General tasks, structured output, tool use

Qwen 2.5 7B Instruct is a well-rounded 7B model from Alibaba’s Qwen series, offering significant improvements over its predecessor in knowledge, coding, math, and instruction following. It supports tool calling, JSON mode, and structured outputs — a feature set that’s rare for models at this price point on Novita AI.

At $0.07 per million tokens for both input and output, it offers flat, predictable pricing. With a 32K context window and support for over 29 languages, it’s a versatile choice for teams that need a capable all-rounder without paying for larger models.

Pros

• Flat $0.07/M pricing for input and output on Novita AI — easy to budget.
• Tool calling, JSON mode, and structured output support.
• 29+ language support with strong multilingual performance.

Cons

• 32K context — shorter than 128K+ models on this list.
• 7B parameters — outperformed by larger models on complex tasks.

Best For

Developers on Novita AI who need a versatile, affordable model with tool use and structured output support for diverse applications.

8. GLM-4.7-Flash

Spec	Detail
Developer	Z.AI
Parameters	~30B
Context Length	200K
Pricing (Input / Output)	$0.07 / $0.40 per M tokens
Quantization	BF16
Best For	Agentic coding, tool use, long-context workflows

GLM-4.7-Flash boasts the longest context window on this list — 200K tokens — and a maximum output of 128K tokens. It’s a 30B-A3B MoE model (30B total, 3B active per forward pass) from Zhipu AI, tailored for agentic coding. It ranks as the strongest model in the 30B class on popular benchmarks like SWE-bench Verified, with strong performance in coding proficiency, long-horizon planning, tool use, and instruction following.

At $0.07/M input tokens on Novita AI, it justifies the cost with full support for tools, JSON mode, structured outputs, reasoning, and a context window that dwarfs everything else here. If you’re building code-generation agents or complex multi-step workflows, this is the cheapest way to get there on Novita AI.

Pros

• 200K context window — largest on this list by far.
• 128K max output — can generate entire codebases in one call.
• Full agentic feature set on Novita AI: tools, reasoning, structured outputs.

Cons

• Output cost ($0.40/M) is steep for heavy generation tasks.
• Input cache pricing ($0.01/M) available for repeated prompts.

Best For

AI coding agents and long-context document analysis on Novita AI that need both thinking and tool use.

9. Qwen3 Coder 30B-A3B

Spec	Detail
Developer	Alibaba (Qwen Team)
Parameters	30.5B (MoE, 3.3B active)
Context Length	160K
Pricing (Input / Output)	$0.07 / $0.27 per M tokens
Quantization	FP8
Best For	Code generation, repo-scale understanding, agentic tool use

Qwen3 Coder 30B-A3B is a 30.5B-parameter MoE model with 3.3B activated weights per forward pass, designed specifically for advanced code generation. It handles repository-scale code understanding, multi-file editing, and agentic tool use with a native context length of up to 256K tokens (160K on Novita AI).

At $0.07 input / $0.27 output per million tokens, it’s the most affordable dedicated coding model on this list. It supports tool calling, JSON mode, and structured outputs — everything you need for building AI-powered development tools.

Pros

• Purpose-built for code with repo-scale understanding.
• 160K context — handles large codebases in a single call.
• MoE efficiency: 30.5B total, but only 3.3B activated weights per call.

Cons

• Specialized for code — may underperform on general conversation tasks.
• Output cost ($0.27/M) higher than general-purpose models.

Best For

Developers on Novita AI building AI coding assistants, automated code review tools, or multi-file code generation pipelines.

10. ERNIE 4.5 21B-A3B

Spec	Detail
Developer	Baidu
Parameters	21B (MoE)
Context Length	120K
Pricing (Input / Output)	$0.07 / $0.28 per M tokens
Quantization	BF16
Best For	Chinese language tasks, cross-modal knowledge, tool use

ERNIE 4.5 21B-A3B is Baidu’s open-source MoE model released under the Apache 2.0 license. It brings an innovative multimodal heterogeneous architecture with improved logical reasoning, mathematical computation, and code generation capabilities. Built on Baidu’s PaddlePaddle framework, it achieves cross-modal knowledge fusion through a parameter-sharing mechanism while maintaining strong performance on Novita AI.

At $0.07 input / $0.28 output per million tokens, it’s competitively priced with tool calling support. It particularly excels at Chinese language tasks, making it an excellent choice for teams serving Chinese-speaking markets through Novita AI.

Pros

• Strong Chinese language performance backed by Baidu’s expertise.
• MoE architecture for efficient inference at $0.07/M on Novita AI.
• 120K context window for long-document processing.

Cons

• Less proven outside Chinese language tasks compared to Llama or Qwen.
• Max output capped at 8K tokens — the lowest on this list.

Best For

Teams on Novita AI targeting Chinese-speaking markets or needing cross-modal knowledge capabilities at an affordable price.

Price Comparison Table

All prices are from Novita AI as of March 2026.

#	Model	Developer	Parameters	Context	Input/M Tokens	Output/M Tokens	Key Strength
1	Llama 3.1 8B Instruct	Meta	8B	16K	$0.02	$0.05	Cheapest general-purpose LLM
2	Qwen3 4B	Alibaba	4B	128K	$0.03	$0.03	Cheapest 128K-context model
3	Llama 3 8B Instruct	Meta	8B	8K	$0.04	$0.04	Flat pricing, proven classic
4	GPT-OSS 20B	OpenAI	21B (MoE)	131K	$0.04	$0.15	OpenAI quality, open-source price
5	Mistral Nemo	Mistral × NVIDIA	12B	60K	$0.04	$0.17	Best multilingual under $0.05
6	GPT-OSS 120B	OpenAI	117B (MoE)	131K	$0.05	$0.25	Most powerful cheap LLM
7	Qwen 2.5 7B Instruct	Alibaba	7B	32K	$0.07	$0.07	Balanced all-rounder, flat pricing
8	GLM-4.7-Flash	Zhipu AI	30B (MoE, 3B active)	200K	$0.07	$0.40	Longest context + agentic coding
9	Qwen3 Coder 30B-A3B	Alibaba	30.5B (MoE, 3.3B active)	160K	$0.07	$0.27	Purpose-built for code
10	ERNIE 4.5 21B-A3B	Baidu	21B (MoE)	120K	$0.07	$0.28	Best for Chinese language

How to Get Started on Novita AI

All 10 models are available through Novita AI‘s API. You can start using any of them in minutes.

Step 1: Get Your API Key

Sign up at Novita AI and grab your API key from the dashboard.

Step 2: Make Your First Call

from openai import OpenAI

client = OpenAI(
    api_key="<Your API Key>",
    base_url="https://api.novita.ai/openai"
)

response = client.chat.completions.create(
    model="meta-llama/llama-3.1-8b-instruct",
    messages=[
        {"role": "system", "content": "You are a helpful assistant."},
        {"role": "user", "content": "Hello, how are you?"}
    ],
    max_tokens=16384,
    temperature=0.7
)

print(response.choices[0].message.content)

Conclusion

The cheapest LLM APIs in 2026 are remarkably capable. For just $0.02 to $0.07 per million input tokens on Novita AI, you get access to models that handle everything from simple chat to advanced reasoning and agentic coding. The days of paying premium prices for production-quality AI are over.

Quick picks on Novita AI:

• Tightest budget? Llama 3.1 8B at $0.02/M — hard to beat.
• Need long context? Qwen3 4B gives you 128K tokens at $0.03/M.
• Need reasoning? GPT-OSS 120B packs 117B parameters into $0.05/M input.
• Need code generation? Qwen3 Coder 30B delivers repo-scale understanding at $0.07/M.

All 10 models are live on Novita AI with APIs, pay-as-you-go pricing, and no rate limits. Sign up, grab a key, and start building.

Novita AI is an AI cloud platform that offers developers an easy way to deploy AI models using our simple API, while also providing an affordable and reliable GPU cloud for building and scaling.

Frequently Asked Questions

Recommended Articles

• Comprehensive Guide to LLM API Pricing: Choose the Best for Your Needs
• Top 6 LLM API for Coding in 2025
• Claude 3 Haiku and Other Budget King LLMs