Meta AI Pricing 2026

Why Meta AI has no price tag

Meta AI is the only assistant among the major consumer platforms with no paid tier at all: no Plus, no Pro, no usage-based upsell. The same assistant is built into Facebook, Instagram, WhatsApp, Messenger, and the standalone meta.ai web and mobile apps, and every one of those surfaces is fully free. That is not a promotional free tier waiting for a paywall; it is the business model. Meta monetizes attention inside its apps, so the assistant's job is engagement and ecosystem lock-in, not subscription revenue. Where OpenAI needs ChatGPT Plus at $19.99/month to fund inference, Meta treats inference cost as a customer-acquisition expense.

The consequences cut both ways. On the plus side: unlimited-feeling everyday chat, image generation, and web-grounded answers without ever entering a card number. On the minus side there is no paid lane to escape rate limiting at peak times, no priority access, no service-level guarantee and no power-user tier with higher caps, because there is no tier above free to sell you. If Meta AI throttles or changes behaviour, the only recourse is another provider.

The Meta AI API in 2026: what changed

Most searches for "Meta AI API pricing" land on outdated pages, because the first-party picture changed materially in mid-2026 and it changed in two directions at once. Meta wound down its first-party Llama API, which had been in public preview since launch, and separately launched a commercial hosted API under its own proprietary models. The open-weights line and the hosted line are now different products from the same company.

On the shutdown, a precise statement of what is and is not established. The wind-down is consistently reported as having taken effect on 6 July 2026, and it is consistent with what the endpoints do now. The first-party deprecation notice that would confirm the date directly no longer resolves: both llama.developer.meta.com and the successor documentation path returned 404 on 2026-08-19. So treat the date as well corroborated rather than cited. The operational conclusion does not depend on the exact day: if a tutorial you are following calls a llama.developer.meta.com endpoint, it is out of date, and that is the reliable tell.

The Llama models themselves are unaffected, because the weights remain openly downloadable. That is why the ecosystem absorbed the shutdown quickly: the same models were already served by third-party inference hosts.

Meta Model API pricing

Meta's current hosted inference is the Meta Model API, serving its Muse Spark models. Read from Meta's own developer documentation at developer.meta.com on 2026-08-19, per million tokens:

Model	Input	Cached input	Output	Context window
Muse Spark 1.1	$1.25	$0.15	$4.25	1M tokens
Muse Spark 1.2	$1.25	$0.15	$4.25	1M tokens
Muse Spark 1.2 Contributor	$0.10	$0.002	$0.20	1M tokens

Two things in that table are worth more than the headline rate. The newer model is priced identically to the one it succeeds, so there is no upgrade premium to budget for, which is unusual enough to be worth noting. And the Contributor variant is an order of magnitude cheaper on every axis: a twelfth the input price, a twentieth the output price and cached input at a fifth of a cent per million tokens. If your workload tolerates it, that line rather than the flagship is the one that decides whether this API is cheap.

New accounts receive a starting credit. No figure is quoted here, because the amount appears only in secondary sources and not on any Meta page that answered on the verification date.

Two traps between the rate card and the invoice

Reasoning tokens bill as output. Muse Spark exposes a reasoning-effort parameter, and every token spent thinking before the model answers is charged at the output rate. A model that thinks for 4,000 tokens before producing a 300-token answer has cost you 4,300 output tokens, not 300. This is the single biggest gap between a headline price and an invoice, and it is not unique to Meta: it applies to every reasoning model with a tunable effort dial, including Qwen's flagship. Benchmark at the effort level you will actually ship, not at the minimum.

A 1M-token context window is a billing surface. Large context invites you to stuff whole codebases and document sets into every request. At $1.25 per million input tokens, a 200,000-token prompt costs $0.25 per call: trivial once, painful at ten thousand calls a day. The $0.15 cached rate is the mitigation, and structuring prompts so that the bulk of them is cacheable is the highest-leverage cost decision available on this API.

What Llama costs on third-party hosts

Because Llama is open-weights, the real API market for Meta's older models is the third-party hosting landscape: AWS Bedrock, Together AI, Groq, Fireworks and DeepInfra all serve Llama models per token. Rates across those hosts fall roughly between $0.05 and $0.90 per million tokens depending on model size and provider, with the smallest variants at the bottom and the largest at the top.

Model	Typical rate per 1M	Notes
Llama 4 Scout	about $0.08 input	The cheapest Meta model by input price
Llama 4 Maverick	$0.15 to $0.63 input	Wide spread by host for one model
Llama 3.3 70B	about $0.10 to $0.59 input	The cheapest and fastest hosts sit at opposite ends
Llama 405B	about $0.80 and up	Widest spread on the board
Self-hosted weights	No per-token cost	You pay compute. Economic only at sustained volume

These are the one part of this page not read from a first-party surface, because there are a dozen first parties. Treat them as a band to plan against and confirm your chosen host's own card before committing.

The number that should change your behaviour is not any single rate, it is the spread: the same Llama model can cost three to twelve times more on one host than another. That is a larger swing than the difference between model sizes. Choosing a smaller model saves you less than choosing the right provider for the model you already picked. Price the host first, and the model second.

Speed is the axis the rate tables hide, and it inverts the ranking. The fastest Llama hosts charge several times what the cheapest ones do and deliver dramatically higher throughput. For an overnight batch job, take the cheap host. For an interactive product where latency is the experience, the expensive host is frequently the cheaper decision once abandoned sessions are counted. Cost per completed task is the metric; cost per million tokens is an input to it.

If you are weighing Llama against the rest of the open-weights field, the open-source model guide maps where it sits against Qwen and DeepSeek, both of which also publish aggressive API pricing (see Qwen pricing and DeepSeek pricing). For routing between providers rather than picking one, OpenRouter alternatives for API routing covers that lane.

A worked example

Rate cards are hard to reason about, so here is one workload priced four ways: a support-triage assistant handling 20,000 requests a month, each with a 3,000-token system prompt and knowledge base, a 500-token user message and a 400-token answer.

Setup	Monthly input	Monthly output	Approximate cost
Muse Spark, no caching	70M at $1.25	8M at $4.25	about $122
Muse Spark, system prompt cached	60M at $0.15, 10M at $1.25	8M at $4.25	about $56
Muse Spark, high reasoning effort	cached as above	about 28M at $4.25	about $141
Llama 4 Scout on a cheap host	70M at about $0.08	8M at about $0.30	about $8

Three conclusions fall straight out. Caching the stable prefix roughly halved the bill without touching quality. Turning up reasoning effort more than undid that saving, because thinking tokens tripled output volume while the rate card never moved. And for a task this mechanical, an open-weights model on a cheap host costs a fraction of the frontier option, which is the argument for routing by task difficulty rather than sending every request to your best model. These are order-of-magnitude figures computed from the rates above.

Is there a free Meta AI API?

Three partial routes, no permanent free lunch. New Meta Model API accounts receive a starting credit, which covers validating a prototype rather than running one. Several third-party Llama hosts offer free tiers or trial credits, where the rate limit is usually the binding constraint rather than the credit balance. And the Llama weights remain free to download, so self-hosting costs compute instead of tokens, which crosses into cheaper only at sustained volume and brings an operations burden with it.

How free Meta AI compares to paid plans

Against the roughly twenty-dollar tier that ChatGPT Plus, Claude Pro and Google AI Pro have converged on, Meta AI's $0 is unbeatable on price. What you give up is documented capability: lighter file handling, no code-execution sandbox and only partial memory, against assistants that publish larger context windows and fuller tooling. The honest framing is that Meta AI is the best $0 assistant, not a $20-class assistant at $0.

The decision rule most people need: if your usage is casual chat, quick lookups and image generation inside apps you already open daily, Meta AI is genuinely sufficient and you should not pay anyone. If you need frontier-model output but not twenty dollars' worth from a single vendor, the middle path is a multi-model plan. Perspective AI at $14.99/month covers models from OpenAI, Anthropic, Google, xAI, DeepSeek and Mistral with mid-conversation switching, below the price of any single official plan. The full cross-vendor math lives in our breakdown of what every AI plan costs, with every vendor's tier sheet in the pricing hub.

When you do not need an API at all

A large share of "Meta AI API pricing" searches come from people who are not building software. They want to use these models, found the API documentation, and are trying to work out what it will cost. If that is you, the API is the wrong product: you would be assembling billing accounts, keys, rate limits and a chat client to reach something a consumer app already does, and the Meta AI assistant itself is free. If the reason you wanted the weights was to keep the transcript on your own hardware, what self-hosting an LLM costs is worked through separately.

How to budget this properly

• Measure cost per completed task, not cost per million tokens. A model at half the rate that needs two attempts and longer prompts is not cheaper. Doubly true for reasoning models, where the effort dial silently moves your output count.
• Design for the cache before you optimise the model. At $1.25 against $0.15, moving stable instructions and reference material into cacheable territory is a larger saving than most model downgrades, and it costs no quality.
• Check the cheap variant first. The Contributor line is an order of magnitude below the flagship on every axis, and plenty of production work never needed the flagship.
• Re-price quarterly. Meta closed a first-party API and launched a commercial one within days of each other in July 2026. This market moves faster than a spreadsheet built two quarters ago, including the figures on this page.

Meta AI pricing FAQ

Will Meta AI add a paid tier?

Meta has announced no consumer subscription for Meta AI, and its ad-funded model gives it little reason to add one: the assistant exists to keep users inside Meta's apps. Nothing prevents a future premium tier, but as of 2026 every Meta AI surface is free and there is no paid upgrade path to buy.

Is the Meta Llama API still available?

No. Meta wound down its first-party Llama API after a public-preview run, reported as taking effect on 6 July 2026, and requests to it no longer succeed. The first-party deprecation notice itself now returns 404, so treat the exact date as corroborated rather than cited. The Llama models remain available: the weights are downloadable, and third-party hosts including AWS Bedrock, Together AI, Groq, Fireworks and DeepInfra continue to serve them.

How much does Llama cost through third-party hosts?

Roughly $0.05 to $0.90 per million tokens, depending on model size and provider. The same model can cost three to twelve times more on one host than another, so compare hosts as carefully as models, and re-verify before committing because this market reprices frequently.