Meta AI API access in 2026: Muse model prices from $0.10 per 1M tokens

Aug 25, 2025

7 min read

Updated: 4 days ago

Updated: November 10, 2026

Developers who want Meta's models through an API now use the Meta Model API, a self-serve, pay-as-you-go service at dev.meta.ai that hosts the Muse family. Prices run from $0.10 per 1M input tokens on Muse Ember up to $4.00 input and $20.00 output on Muse Wonder, with no free tier, credits or batch discount listed on the official pricing page. Llama models are not served through this hosted API, and Meta's developer help center directs Llama users to download the models from llama.com and host them themselves.

The service works with the OpenAI SDK and the Anthropic SDK, exposes three request protocols on one base URL, and gives its three text models a context window of 1,048,576 tokens. The tables below list every hosted model, its official price and its rate limit, followed by the setup steps and a monthly cost example computed for a high-volume workload.

........

Model (API ID)	Input / cached input / output per 1M tokens	Context window	Main use
Muse Ember (muse-ember-1.0)	$0.10 / $0.004 / $0.50 up to 200K input	1,048,576 tokens	High-volume sub-agent tasks, routing, tool calls
Muse Spark 1.3, 1.2, 1.1 (Standard)	$1.25 / $0.15 / $4.25	1,048,576 tokens	General model, Meta's recommended default
Muse Spark 1.3 and 1.2 (Contributor)	$0.10 / $0.002 / $0.20	1,048,576 tokens	Discounted, content may train Meta models
Muse Wonder (muse-wonder-1.0)	$4.00 / $0.20 / $20.00	1,048,576 tokens	Hardest reasoning and planning, in preview
Muse Image (muse-image-1.0)	$0.01 per generated image	Not stated	Image generation and editing
Muse Voice Transcribe	$0.18 per hour of audio	Not applicable	Speech to text
SAM 3.1 (sam-3.1)	$2.50 per 1,000 images, $0.20 per 1,000 video frames	Not stated	Segmentation boxes and masks

........

··········

META MODEL API: HOW DEVELOPER ACCESS TO META AI WORKS IN 2026.

The Meta Model API is the hosted service through which Meta sells access to its current models. Meta describes it as self-serve, with a one-click API key created from the dashboard at dev.meta.ai, and it bills per token or per unit with no subscription required. Requests go to the base URL https://api.meta.ai/v1 with a bearer token, and the documentation names the environment variable MODEL_API_KEY in its examples.

Three protocols run on that base URL. The Responses API is Meta's choice for agentic, multi-step work with server-managed state and reasoning replay; the Chat Completions API follows the OpenAI messages format; and the Messages API mirrors Anthropic's format, so code written for either SDK can be pointed at Meta by changing the base URL, the key and the model ID. Meta's documentation also mentions agent frameworks such as the Claude Agent SDK and the OpenAI Codex app-server.

The Llama family sits outside this service. Meta's developer help center states that Llama models are not run for customers through Model API and points to llama.com for downloading and self-hosting, and the old Llama API documentation address now redirects to the Model API overview. Teams that built on Llama endpoints during 2025 therefore have two paths, self-hosting a Llama model or moving to a Muse model on the hosted API.

··········

⁣⁣⁣⁣⁣⁣⁣⁣⁣⁣⁣⁣

MUSE EMBER, MUSE SPARK AND MUSE WONDER: API PRICES AND LIMITS PER MODEL.

Muse Ember is the smallest and cheapest text model on the API. Meta calls it the efficient execution model of the Muse family and recommends it for sub-agent coding tasks, terminal commands, lookups, routing and running one task across large batches of files or records. Its model page lists a 1,048,576-token context window, up to 524,288 output tokens, input in text, image, video, audio and PDF, and reasoning effort settings from "minimal" to "max", with "max" as the default. Ember is the only model with a long-context price: when a request carries more than 200K input tokens, counting cached tokens and any stored conversation pulled in with previous_response_id, the whole request is billed at $0.30 input, $0.012 cached input and $1.50 output per 1M tokens. Our report on Muse Ember's API launch and benchmarks covers the launch figures in detail.

Muse Spark is available in three versions, with muse-spark-1.3 used in Meta's quickstart as the default. In Standard mode it costs $1.25 per 1M input tokens, $0.15 cached and $4.25 output. Versions 1.3 and 1.2 are also sold as Contributor variants priced at $0.10 input, $0.002 cached and $0.20 output, in exchange for permission to use the submitted content to train future Meta models.

Muse Wonder is Meta's most capable API model and the most expensive, at $4.00 input, $0.20 cached and $20.00 output per 1M tokens. The models page marks it as a preview. The stronger Muse Wonder Full is not on the API at all; Meta sells it only inside Muse Code subscriptions.

........

Item	Muse Ember	Muse Spark (Standard)	Muse Spark (Contributor)	Muse Wonder
Input per 1M tokens	$0.10 (over 200K: $0.30)	$1.25	$0.10	$4.00
Cached input per 1M tokens	$0.004 (over 200K: $0.012)	$0.15	$0.002	$0.20
Output per 1M tokens	$0.50 (over 200K: $1.50)	$4.25	$0.20	$20.00
Requests per minute	3,000	3,000	150	3,000
Tokens per minute	4,000,000	4,000,000	3,000,000	4,000,000
Content used for training	No	No	Yes	No

........

Web search grounding through the web_search tool adds $2.50 per 1,000 search queries on Muse Wonder, Muse Spark and Muse Ember, on top of token costs. System-prompt tokens that Meta injects are neither billed nor counted in usage.

··········

META MODEL API RATE LIMITS, CACHING AND DATA RETENTION RULES.

Rate limits apply per team and per model, not per API key, and the pricing page lists them as fixed values with no usage tiers to climb. Muse Image is limited to 150 requests per minute, SAM 3.1 to 3,000, and Muse Voice Transcribe to 256 concurrent streams and 16,000 streams per hour. Background responses submitted with background set to true have a separate cap of 600 new submissions per minute per team. Over the limit, the API returns HTTP 429 with a Retry-After header, and Meta recommends exponential backoff with jitter starting around 500 milliseconds.

Prompt caching is automatic, with no flag to set. Developers can send a prompt_cache_key to raise hit rates at high volume, and on the Responses API the setting prompt_cache_retention set to "24h" asks for retention of up to 24 hours, which Meta describes as a hint the server may not honor under load. Cached tokens appear in usage.input_tokens_details.cached_tokens on the Responses API and in usage.prompt_tokens_details.cached_tokens on Chat Completions. A token-counting endpoint, POST /v1/responses/input_tokens, checks whether an input fits the context window without generating output.

Content sent in Standard mode is not used to train Meta models, according to the pricing page. Zero data retention is offered to qualified organization accounts only, enabled by Meta after a request through the sales contact form; personal accounts are not covered, ZDR support may vary by model, and flagged policy violations can be kept for up to two years even under ZDR. Model availability also varies by country under Meta's Geographic Use Policy.

··········

[[ADPLUS_300x250_1]]

HOW TO GET A META MODEL API KEY AND MAKE THE FIRST CALL.

The official quickstart takes four steps: open the dashboard at dev.meta.ai, go to API keys and create a key, store it in the MODEL_API_KEY environment variable, and call the API with an SDK or plain HTTP. The OpenAI SDK does not read MODEL_API_KEY by itself, so the key has to be passed explicitly. This Python example is taken from Meta's quickstart:

import os

from openai import OpenAI

client = OpenAI(
    base_url="https://api.meta.ai/v1",
    api_key=os.environ["MODEL_API_KEY"],
)

response = client.responses.create(
    model="muse-spark-1.3",
    input="What is the capital of France?",
)

print(response.model_dump_json(indent=2))

A 401 invalid_api_key error means the key is missing or wrong, and a 404 model_not_found error means the model ID is misspelled. Swapping muse-spark-1.3 for muse-ember-1.0 or muse-wonder-1.0 is enough to move the same code to another model.

··········

MUSE CODE SUBSCRIPTIONS VERSUS PAY-AS-YOU-GO API BILLING.

Muse Code is Meta's coding agent, run from a command-line tool, and it can be paid in two ways. With a Model API key it bills per token at the API prices above; with a subscription it draws on plan allowances instead. Meta's subscription page lists four plans, Everyday Usage, High Usage, Power Usage and Maximum Usage, with allowances of 10 to 50 prompts every five hours on the first plan and 5x, 20x and 80x that allowance on the higher plans. Muse Wonder Full is included only from Power Usage upward. Meta's documentation does not publish the monthly prices of these plans; they appear at checkout. Subscriptions work only through the Muse Code CLI, and any extra API key beyond the one connected to Muse Code is billed pay-as-you-go. More background on the agent and the coding model behind it is in our coverage of Muse Spark 1.3 and Muse Code.

··········

⁣⁣⁣⁣⁣⁣⁣⁣⁣⁣⁣⁤

META MODEL API COST EXAMPLE: 300,000 REQUESTS A MONTH ON EACH MUSE MODEL.

To compare the models on a realistic workload, take a routing or support service that handles 10,000 requests a day for 30 days. Each request sends 3,000 input tokens, of which 2,000 are a repeated system prompt served from cache, and returns 400 output tokens. All requests stay under 200K input tokens, so Muse Ember is billed at its standard rates. The arithmetic below was run in Python with the official prices.

........

Model	Uncached input cost	Cached input cost	Output cost	Monthly total
Muse Ember	$30.00	$2.40	$60.00	$92.40
Muse Spark (Contributor)	$30.00	$1.20	$24.00	$55.20
Muse Spark (Standard)	$375.00	$90.00	$510.00	$975.00
Muse Wonder	$1,200.00	$120.00	$2,400.00	$3,720.00

........

The Contributor variant of Muse Spark is the cheapest line, but only for teams that accept Meta training on their content; its limit of 150 requests per minute still covers this workload's average of about 7 requests per minute, with room for traffic peaks. Among the models that keep content out of training, Muse Ember costs about one tenth of Muse Spark Standard on this workload and about 2.5% of Muse Wonder. The long-context rule changes the picture for large documents: a single Ember request with 190K input tokens and 5K output tokens costs about $0.02, while the same request at 250K input tokens moves entirely to the higher tier and costs about $0.08.

··········

META AI API IN 2026: MODELS, PRICES AND OPEN QUESTIONS.

Meta's hosted API in 2026 is a Muse-only service with published per-token prices, fixed per-model rate limits, OpenAI and Anthropic SDK compatibility and a 1M-token context window on its three text models. Muse Ember sets the floor at $0.10 per 1M input tokens with full training opt-out, Muse Spark Standard is the general default at $1.25, and Muse Wonder at $4.00 and $20.00 is reserved for the hardest work while still in preview.

Several points remain undocumented on Meta's official pages. The models page does not state a release status for any model except Muse Wonder, nor deprecation dates, and the Muse Code subscription prices are shown only at checkout. Meta's research blog post on Muse Ember did not load when we checked, so the figures above come from the developer documentation and pricing page. Self-hosted open-weight options, including Muse Spark Open and the Llama family, are billed by whichever infrastructure provider runs them, not by Meta.

··········

·····

DATA STUDIOS

·····

datastudios.org