There is no single “Llama API” to sign up for. Meta releases Llama as open weights, and Meta’s own developer platform no longer lists Llama models in its API catalog, so in practice you reach Llama through a third-party host — Novita AI, DeepInfra, or Groq today. Which version you pick (Llama 3.3 70B, Llama 4 Scout, or Llama 4 Maverick) sets your usable context window anywhere from 12K to more than 1 million tokens, and which host you pick decides whether you get a published per-token rate at all. The table below puts every host-and-version combination that is actually callable side by side.
Hosted Llama API options at a glance
| Provider | Llama version | Input ($/1M tokens) | Output ($/1M tokens) | Context window |
|---|---|---|---|---|
| Novita AI | Llama 3.3 70B | $0.135 | $0.40 | 12K |
| Novita AI | Llama 4 Scout | $0.18 | $0.59 | 128K |
| Novita AI | Llama 4 Maverick | $0.27 | $0.85 | 1M |
| DeepInfra | Llama 3.3 70B | $0.10 | $0.32 | 128K |
| DeepInfra | Llama 4 Scout | $0.10 | $0.30 | 320K |
| Groq | Llama 3.3 70B | Contact sales | Contact sales | 131K |
| Groq | Llama 4 Scout / Llama 4 Maverick | Contact sales | Contact sales | Not published |
| Together AI | No Llama model in the current serverless catalog | — | — | — |
| Meta (first-party API) | Not offered | — | — | — |
Pricing and context windows come from each provider’s own current model or pricing catalog, checked October 16, 2026. Groq documents its Llama models on an enterprise tier priced by contract, with no published per-token rate; its API model list also includes Llama 4 Scout and Llama 4 Maverick. Together AI’s serverless catalog lists no Llama model, and Meta’s first-party developer platform does not list Llama at all.
Three patterns stand out. Novita AI is the only host in this set that serves all three versions with published per-token pricing. Llama 4 Maverick is the least widely hosted version — the hosts that publish a rate for Scout do not all add Maverick. And only two hosts here publish a per-token rate at all, Novita AI and DeepInfra, while Groq prices its Llama models by contract, so a like-for-like sticker comparison is not always available.
Why “Llama API” means a host, not a vendor
Meta publishes Llama as open weights under the Llama license, so anyone can download the weights and run them on their own hardware. That is the origin of the confusion around “the Llama API”: there is no vendor-operated endpoint that is the one canonical way to call Llama.
Meta did preview a first-party hosted Llama API, and the official Python SDK for it still exists on GitHub and PyPI. But Meta’s current developer platform documents a different catalog — the Meta Model API, which serves Meta’s Muse models (Muse Wonder, Muse Spark, Muse Image, Muse Voice Transcribe) plus the Segment Anything model and the open-weight Muse Glimmer. No Llama model appears in that catalog, and the SDK’s last release was version 0.6.0 in December 2025. If you want Llama behind an API today, plan on a third-party host rather than Meta directly.
That leaves two practical routes:
- Hosted API. A provider runs the weights and exposes an OpenAI-compatible chat completions endpoint. You pay per token, get elastic capacity, and never touch a GPU.
- Self-hosting. You rent GPUs and serve the weights yourself. This makes sense when you need data residency, custom quantization, or a fine-tune that a host does not offer — but it moves cost from per-token billing to per-GPU-hour billing plus the engineering to keep the server up.
This guide covers the hosted route, because that is what “Llama API” almost always means in practice.
Llama 3.3 70B vs Llama 4 Scout vs Llama 4 Maverick
The three versions are not generations of the same design. Llama 3.3 70B is a dense text model; the Llama 4 family is built on a mixture-of-experts design and takes images as well as text.
| Version | Design | Input | Native context | Typical fit |
|---|---|---|---|---|
| Llama 3.3 70B | Dense, 70B parameters | Text | 128K | General chat, summarization, and code where cost per token matters and images are not needed |
| Llama 4 Scout | MoE, 17B active parameters, 16 experts | Text and images | Up to 10M | Long documents and image-plus-text prompts at the lowest Llama 4 price |
| Llama 4 Maverick | MoE, 17B active parameters, 128 experts | Text and images | Up to 1M | The strongest Llama 4 quality, and the version to reach for when one prompt spans a very long context |
Native context and served context are different numbers. Meta designs Scout for up to 10 million tokens, but a host decides how much of that window it exposes: Novita AI serves Scout at 128K and DeepInfra at 320K. Always check the host’s own model page for the window you actually get, not the maximum the architecture allows.
How the hosts compare
Novita AI is the only host here that carries all three versions with public pricing, and it exposes them through one OpenAI-compatible endpoint at https://api.novita.ai/openai. Llama 3.3 70B runs at $0.135 input / $0.40 output per million tokens, Llama 4 Scout at $0.18 / $0.59, and Llama 4 Maverick at $0.27 / $0.85. The window is the one number to check per version: Novita serves Llama 3.3 70B with a 12K window, so match it to your prompt length, while Llama 4 Scout and Llama 4 Maverick come with much larger windows on the same endpoint.
DeepInfra publishes per-token rates for Llama 3.3 70B ($0.10 / $0.32) and Llama 4 Scout ($0.10 / $0.30), and serves Scout with a 320K window. It does not list Llama 4 Maverick, so its Llama lineup stops at two of the three versions.
Together AI does not currently list a Llama model in its serverless catalog. Its published serverless chat models are DeepSeek, Qwen, Kimi, GLM, GPT-OSS, and Meta’s newer Muse models, with no Llama entry. If you standardize on Together for other model families, treat Llama as a separate integration rather than an assumption, and check the catalog again before you plan around it.
Groq built its reputation on fast Llama inference. Its model documentation places the Llama models it shows — Llama 3.1 8B and Llama 3.3 70B — on an enterprise tier priced by contract rather than a published per-token rate, and its API model list also includes Llama 4 Scout and Llama 4 Maverick. Because every Llama model there is contract-priced, you cannot compare Groq on sticker price for any version; if low latency is the deciding factor, it is worth a sales conversation rather than a line in a price table.
How to choose by scenario
Budget first. Llama 3.3 70B is the cheapest way to run a capable text model, and Novita AI serves it at $0.135 / $0.40 on the same OpenAI-compatible endpoint as its Llama 4 models. The window is the tradeoff to check: Novita AI serves 3.3 70B at 12K, so if your prompts fit within that, the smaller window costs you nothing and you keep one integration across the whole Llama lineup.
Long context first. Llama 4 Maverick on Novita AI is the only listing here that reaches a full 1M tokens, which is the version to pick when a single prompt spans an entire codebase or a long transcript. If your longest prompts stay within 128K, Llama 4 Scout on Novita AI handles them on the same endpoint.
Multimodal first. Llama 3.3 70B is text-only, so any image-plus-text workload starts at Llama 4 Scout. Novita AI serves Scout at $0.18 / $0.59 and Maverick at $0.27 / $0.85, both accepting images through the same endpoint, so you can move up to Maverick for higher quality without changing the integration.
Latency first. Groq is the host built around fast Llama inference, but its Llama lineup is contract-priced, so it cannot be ranked against the published rates here and needs a sales conversation. If you cannot commit to a contract, benchmark the hosts that publish rates on your own prompts instead.
One vendor for more than inference. If your stack also needs GPU capacity or sandboxed code execution, consolidating on a host that sells those alongside model calls avoids a second billing relationship. Novita AI is the one in this comparison that bundles LLM inference with GPU Cloud and Agent Sandbox.
Calling a Llama model on Novita AI
Every Llama model on Novita AI is reachable through the OpenAI-compatible endpoint, so existing OpenAI SDK code works after a base URL, API key, and model ID change:
from openai import OpenAI
client = OpenAI(
base_url="https://api.novita.ai/openai",
api_key="YOUR_NOVITA_API_KEY",
)
response = client.chat.completions.create(
model="meta-llama/llama-4-scout-17b-16e-instruct",
messages=[{"role": "user", "content": "Summarize what Llama 4 Scout is best at in two sentences."}],
max_tokens=256,
)
print(response.choices[0].message.content)Swap the model value to meta-llama/llama-3.3-70b-instruct or meta-llama/llama-4-maverick-17b-128e-instruct-fp8 to call the other versions on the same endpoint. A raw HTTP call looks like this:
curl https://api.novita.ai/openai/v1/chat/completions \
-H "Content-Type: application/json" \
-H "Authorization: Bearer $NOVITA_API_KEY" \
-d '{
"model": "meta-llama/llama-4-maverick-17b-128e-instruct-fp8",
"messages": [{"role": "user", "content": "Hello"}],
"max_tokens": 512
}'Model IDs are host-specific. The same Llama 4 Scout weights are named differently on different providers, so keep the ID in configuration rather than scattered through code.
Common pitfalls
Assuming every host carries every version. Llama 4 Maverick is the clearest example: several hosts that added Scout never added Maverick. Check availability per version, not per brand.
Reading native context as served context. Meta’s architecture figures (10M for Scout, 1M for Maverick) are the maximums the design supports. What you can send depends on the host, and the host’s number is the one that limits your prompt.
Ignoring the 3.3 70B context gap. The cheapest Llama 3.3 70B listing is not automatically the best value: a 12K served window and a 128K served window are different products for a long-prompt workload, even at a similar price.
Expecting a first-party option. Meta’s own API no longer serves Llama, so budgeting for “the official Llama endpoint” is a dead end. Plan around a third-party host from the start.
Forgetting that pricing moves. Per-token rates and served context windows on every host here change independently of this comparison. Re-check the provider’s own page before you size a production budget.
Conclusion
“Llama API” is not one endpoint — it is a set of open-weight models plus a handful of hosts that serve them, and the right pick depends on which version you need and what you are optimizing for. Novita AI is the only host in this comparison that serves Llama 3.3 70B, Llama 4 Scout, and Llama 4 Maverick together with public pricing, through one OpenAI-compatible endpoint and alongside GPU Cloud and Agent Sandbox; DeepInfra publishes rates for 3.3 70B and Scout and the widest Scout window but carries no Maverick; Groq serves Llama 4 Scout and Maverick as well as the 3.3 generation but prices all of them by contract; Together AI’s current serverless catalog lists no Llama model; and Meta’s first-party platform no longer lists Llama at all. Start from the version that matches your workload — text-only and cheap, multimodal and long-context, or maximum quality — then pick the host whose lineup, price, and served context window fit.
FAQ
Is there an official Llama API from Meta?
Not in Meta’s current developer catalog. Meta releases Llama as open weights, and its developer platform now documents the Meta Model API, which serves Meta’s Muse models and Segment Anything rather than Llama. The official Llama API Python SDK still exists, but its last release was in December 2025. For API access to Llama today, use a third-party host.
What is the difference between Llama 3.3 70B and Llama 4?
Llama 3.3 70B is a dense, text-only model. The Llama 4 family uses a mixture-of-experts design and accepts images as well as text: Llama 4 Scout is the lower-cost, very-long-context option, and Llama 4 Maverick is the higher-quality one. If you do not need images and your prompts are short, Llama 3.3 70B is usually the cheaper choice.
Which host is cheapest for Llama?
What counts as “cheapest” depends on which version you need, not on a single per-token number. Novita AI is the only host here that serves Llama 3.3 70B, Llama 4 Scout, and Llama 4 Maverick together with published pricing, so one integration covers the full lineup and you are not adding a second provider when you move between versions. The version you actually need matters more than a few cents of per-token difference — and so does the served context window, which differs between hosts even for the same model.
Does the native context window apply when I call a host?
No. Meta’s architecture figures are maximums the model design supports. Each host exposes its own window, and that is the number that limits your request. Llama 4 Scout is designed for up to 10M tokens natively, but the hosts here serve it at 128K or 320K.
Can I use my existing OpenAI SDK code to call Llama?
Yes, on hosts that expose an OpenAI-compatible endpoint. On Novita AI you point the OpenAI SDK at https://api.novita.ai/openai, set your Novita API key, and set model to the host’s model ID. The request and response shapes match, so only the connection details and the model name change.
Sources
- Novita AI model pages — Llama 3.3 70B, Llama 4 Scout, and Llama 4 Maverick pricing and context windows; checked October 16, 2026.
- Meta Model API documentation — current Meta-hosted model catalog; checked October 16, 2026.
- Meta Llama 4 announcement — Llama 4 Scout and Maverick architecture and native context; checked October 16, 2026.
