A8gent
AI Model ProvidersOpen weights

Meta Llama API Pricing for Small Business Automation

Not a vendor so much as a licence and a set of weights, hosted cheaply by everyone and runnable on your own hardware.

Researched by A8gent
Contents
  1. Verdict
  2. API pricing
  3. Llama is a licence and some weights, not a vendor
  4. What Llama costs, and why the host matters more than the model
  5. Reliability when the model and the vendor are different things
  6. What self-hosting Llama actually takes
  7. Where Llama fits in a small business stack
  8. How to access
  9. Privacy and residency
  10. What we build
  11. FAQ
  12. Price sources
VerdictShould you run automations on Meta Llama?

Llama is best understood as infrastructure rather than a product. Meta does not really sell you an API any more; it publishes weights under its own licence and a competitive market of hosts sells inference on top. That market is why the small models are among the cheapest capable options anywhere, with Llama 3.1 8B at around $0.05 in and $0.08 out per 1M tokens on Groq as of August 2026. For a small business the honest verdict is that Llama earns its place when you need very cheap high-volume inference, very low latency, or the model running inside your own network. If you need none of those, a mainstream API is less work.

Best for

Very high-volume simple tasks, latency-sensitive agent loops on fast inference hosts, and on-premise deployments where data cannot leave your network.

Not for

Businesses that want one vendor, one bill and one support contact, since the Llama route means picking a host and owning the integration differences.

Meta Llama API pricing per 1M tokens

There is no single Llama price because there is no single Llama vendor. Meta publishes weights and a market of hosts competes on price, latency and features. That competition is the reason Llama inference is cheap, and it is also why shopping around is not optional. Llama 3.3 70B, the same weights, listed at $0.10 per 1M input on OpenRouter, $0.59 on Groq and $1.04 on Together AI when we checked in August 2026. That is a ten-fold spread on an identical model, and the expensive ones are not overcharging: Groq's documented 280 tokens per second on that model is what you are paying for. Decide whether you are buying tokens or speed, then pick accordingly. One more note on lineup: the newest generation we could verify at any host is Llama 4, and we found nothing called Llama 5 on Groq, Together AI or OpenRouter's Meta Llama catalogue. Prices below are as of August 2026.

Meta Llama API pricing per 1M tokens as of August 2026
ModelContextInput / 1MOutput / 1M
Llama 3.1 8B Instant (Groq)Documented at 560 tokens per second on Groq's model page131,072$0.05$0.08
Llama 3.1 8B Instruct (OpenRouter)Same weights, less than half Groq's price, without Groq's speed131K$0.02$0.04
Llama 3.3 70B Versatile (Groq)Documented at 280 tokens per second. On 17 Nov 2026 Groq's own docs listed this model as contact sales; $0.59 / $0.79 is still Groq's rate on OpenRouter131,072$0.59$0.79
Llama 3.3 70B Instruct (OpenRouter)Roughly a sixth of Groq's input price for identical weights131K$0.10$0.32
Llama 3.3 70B (Together AI)Ten times OpenRouter's input rate for the same model. On 17 Nov 2026 Together listed it as a dedicated deployment, not yet liveNot published on the pricing page$1.04$1.04
Llama 4 Scout (OpenRouter)Mixture-of-experts with 16 experts1.31M$0.10$0.30
Llama 4 Maverick (OpenRouter)Mixture-of-experts with 128 experts, multimodal1.05M$0.20$0.696
Llama 3.2 1B Instruct (OpenRouter)The cheapest input rate in this entire directory60K$0.027$0.201
Llama Guard 4 12B (OpenRouter)Safety classifier, useful as a guardrails layer in front of another model1.05M$0.18$0.18
Llama 5We found no Llama 5 listed on Groq, Together AI or OpenRouter's Meta Llama catalogue in August 2026, and could not verify hosted pricing anywhereNot publishedNot publishedNot published

List prices as of August 2026, taken from the provider's published pricing page on August 5, 2026. Model prices move fast - check the source links before you budget.

Llama is a licence and some weights, not a vendor

The most common confusion about Llama is treating it like OpenAI or Anthropic - a company you sign up with and send requests to. It is not that. Meta trains models, publishes the weights under its own licence, and by 2026 appears to have wound down its first-party paid API, directing developers to third-party hosts instead. What you actually buy is inference from a host who has loaded those weights onto their hardware.

That structure explains almost everything about Llama's economics. Because hosts are selling essentially the same product, they compete on price, speed and features rather than on model exclusivity, and inference prices fall to something close to the cost of the hardware. An 8B-class model at around $0.05 per million input tokens on a fast host is not a promotional rate, it is what commodity inference costs.

It also explains the friction. There is no single support contact, no single data policy, no single set of rate limits, and no vendor accountable end to end. Tool-calling behaviour genuinely varies between hosts running identical weights, because each implements its own function-calling parser and schema handling. An agent that works on one host can misbehave on another, which is a failure mode that does not exist with a single-vendor API.

The practical mitigation is to constrain output rather than trust the host. Use grammar-constrained or strict-JSON output where the host supports it, validate every tool call against your schema before executing it, and include the host in your evaluation runs rather than only the model. Do that and the portability becomes an asset instead of a hazard.

What Llama costs, and why the host matters more than the model

We price every provider on this site against the same two workloads. Workload A is an inbox agent: 1,500 runs a month at about 3,000 input tokens each plus 400 output. Workload B is document extraction: 4,000 documents a month at 6,000 input and 300 output tokens each. On Llama 3.1 8B via Groq at $0.05 and $0.08, that is about $0.27 and $1.30 a month. On the same weights via OpenRouter at $0.02 and $0.04, about $0.11 and $0.53. Those are the cheapest numbers anywhere in this directory.

Step up to Llama 3.3 70B and the host choice becomes the whole decision. Workload B costs about $2.78 a month on OpenRouter at $0.10 and $0.32, about $15.11 on Groq at $0.59 and $0.79, and about $26.21 on Together AI at $1.04 flat. Same weights, same task, nearly ten times the bill. Nobody is being dishonest here: Groq documents 280 tokens per second on that model against 560 on the 8B, and you are paying for hardware built to deliver it. But if your job runs overnight and nobody is waiting, you are paying a large premium for speed you will never use.

The Llama 4 tier is where the value sits for anything needing a large context. Llama 4 Scout lists at $0.10 and $0.30 with a 1.31M context on OpenRouter, and Llama 4 Maverick at $0.20 and $0.696 with 1.05M and multimodal input. Workload B on Scout is about $2.76 a month. For comparison, the same job is about $6.24 on OpenAI's budget tier and about $30.00 on Claude Haiku 4.5.

One structural warning that no price list shows. There is no prompt caching discount to reach for and no batch API in the sense that OpenAI or Anthropic mean it, because those are vendor features rather than model features and each host decides independently what it offers. When you compare a Llama host against a mainstream API, compare the mainstream API's cached and batched rate, not its list rate, or you will overstate the saving.

Reliability when the model and the vendor are different things

Every reliability concern that applies to other providers applies to Llama, plus one that is unique to it: the same weights behave differently depending on who is serving them. Each host implements its own function-calling parser, schema handling and stop-token logic on top of identical model files. An agent that reliably emits well-formed tool calls on one host can produce subtly malformed ones on another, and no benchmark of the model will warn you, because the model is not the variable.

The mitigation is to stop trusting the host and start constraining the output. Use grammar-constrained or strict JSON output where the host supports it, validate every tool call against your schema before you execute it, and treat a failed validation as a retryable error rather than a crash. Include the host in your eval runs, not just the model, and re-run the eval when you switch hosts even if the model name is unchanged.

The provider-neutral research applies too. LOCA-bench, published in early 2026, found that most models including open-source ones degrade sharply as context grows, and explicitly notes the gap between models widening at longer contexts, which is a caution against using Llama 4 Scout's 1.31M window as an operating target rather than a ceiling. AgentCE-Bench, from April 2026, found a 10 percent tool-call failure rate visibly degrading agent performance at every horizon tested, which is exactly the failure rate a mismatched host parser can introduce on its own.

The compensating advantage is real and underrated: no vendor lock-in. If a host raises prices, degrades, or changes terms, you move. The model your automation was tested against is the same model somewhere else. OpenRouter's automatic failover across hosts turns that portability into an availability strategy, which is something no single-vendor API can offer at any price.

What self-hosting Llama actually takes

Llama is the family most likely to make on-premise automation realistic for a small business, because the small models are genuinely small. As a rough guide, an 8B-class model needs about 16GB of VRAM at full precision and roughly 6 to 8GB quantised to four bits, which is a single consumer graphics card. A 70B-class model needs about 140GB at full precision or roughly 40GB quantised, meaning one or two data-centre cards. The very large frontier models are multi-node deployments and out of scope for a small business.

Those are standard community rules of thumb rather than figures from a Meta document, so treat them as planning estimates. The important point is the shape: a useful automation model can run on hardware you can buy once and put in a cupboard, which is not true of most frontier open weights.

Whether to do it comes down to why. If the driver is regulatory or contractual - the data legally cannot leave your premises - self-hosting is the only route and the cost is the price of compliance. If the driver is saving money, do the arithmetic first: hardware, electricity, and above all the engineering time to keep it patched, monitored and available will usually exceed a few hundred dollars a month of API spend. API spend is someone else's 3am problem, and that is worth paying for.

The middle option that suits most small businesses is a fine-tuned small model on a managed host. You get the cost profile of a small model, the accuracy of one trained on your specific task, and none of the operational burden. For a single repetitive job - routing tickets into your categories, extracting your fields from your forms - a fine-tuned 8B model frequently beats a general frontier model on both cost and consistency.

Where Llama fits in a small business stack

Three situations make Llama the right call. The first is volume so high that per-token cost genuinely dominates: monitoring every message, scoring every record, checking every line. At five cents a million input tokens, cost stops being a design constraint entirely.

The second is latency. Specialist hosts running Llama on purpose-built hardware deliver token throughput that the frontier APIs do not match, and in an agent loop that makes ten sequential calls, per-call latency multiplies into a wait a human notices. If someone is sitting watching a spinner, this matters more than benchmark scores.

The third is on-premise. If the data cannot leave your network, no API vendor's contractual assurance solves it and self-hosting a small Llama model does.

Outside those three, we would usually point a small business at a mainstream API instead. The reason is not capability, it is total effort: one vendor, one bill, one support contact, documented behaviour and a first-class integration in whatever automation platform you are using. The cheapest tokens in the world are poor value if the build takes three times as long and nobody is accountable when it breaks. If you want help working out which side of that line your workflow falls on, that is precisely what a free automation audit covers.

Strengths for agent workloads

  • The cheapest hosted inference of any capable model family, because hosts compete on commodity weights
  • Extremely fast inference available from specialist hosts, which matters when an agent makes many sequential calls
  • Genuine self-hosting: the small models run on a single consumer GPU, so a fully on-premise automation is realistic
  • No vendor lock-in at the model layer, since the same weights are available from many hosts and you can move
  • Wide fine-tuning support across hosts, which is practical for narrow repetitive tasks

Weaknesses

  • The Llama Community Licence is not an OSI open-source licence and carries conditions, including a very large user threshold and an acceptable-use policy, that need legal review at scale
  • Tool-calling behaviour varies by host, not just by model, because each host implements its own function-calling parser
  • Meta's own first-party paid API appears to have been wound down, so there is no single vendor relationship to fall back on, and llama.com now redirects to Meta's general developer site
  • Documentation is fragmented across Meta and many hosts, and we could not verify a Llama 5 generation at any host as of August 2026
  • Host catalogues are thinning: Together AI's serverless chat pricing table listed only two Llama models when we checked, which is a signal about where hosts are putting capacity
  • Price dispersion between hosts is enormous, up to ten times on identical weights, so a casual host choice can cost you an order of magnitude
  • Frontier capability lags the top closed models, so the hardest reasoning steps are usually better placed elsewhere

How to access Meta Llama

  • Specialist inference hosts such as Groq, Together AI, Fireworks and Cerebras. This is the normal route. They differ substantially on price, latency, model selection and fine-tuning support, so compare rather than defaulting.
  • AWS Bedrock and other cloud marketplaces, which is the easiest procurement path if you already buy through a hyperscaler and want the model under an existing contract and region.
  • OpenRouter, which routes across hosts with automatic failover, useful for reliability and for testing hosts against each other without separate integrations.
  • Self-hosting from Hugging Face. An 8B-class model needs roughly 16GB of VRAM at full precision or about 6 to 8GB quantised, which is a single consumer card. A 70B-class model needs roughly 140GB at full precision or about 40GB quantised, so one or two data-centre cards. These are standard rules of thumb rather than Meta figures.

Data privacy and residency

  • There is no single Llama data policy, because your data goes to whichever host you chose. Read that host's policy, not Meta's. Groq, Together, Fireworks, Bedrock and OpenRouter all publish their own retention and training terms and they are not the same.
  • This fragmentation cuts both ways. It means one more thing to check, and it also means you can pick a host whose terms suit you, or move if they change, without changing the model your automation was tested against. That portability is a real and underrated advantage.
  • Self-hosting is the strongest privacy position available to any small business: the weights run on your hardware and no prompt leaves your network. For an 8B-class model on a single GPU this is genuinely achievable, which is not true of most open flagships.
  • The licence deserves attention separately from privacy. The Llama Community Licence is frequently described as open source and is not; it is a custom licence with conditions including a very large monthly-user threshold and an acceptable-use policy. For an ordinary small business it is unlikely to bite, but if you are embedding Llama in a product you sell, have someone read it.

What we build on Meta Llama

  • Very high-volume simple classification and tagging where the cost per run has to be effectively zero
  • Latency-sensitive agent loops on a fast inference host, where each sequential call adds to a user's wait
  • Fully on-premise automation on a single GPU, for businesses whose data cannot leave the building
  • Fine-tuned narrow models for one repetitive task, where a small tuned model beats a large general one on both cost and consistency
  • The volume tier of a two-model stack, with a frontier model handling the judgement steps

Model choice is a line item, not a strategy. If you want the workflow designed, built and handed over with a plain-language report of what it ran and what it saved, start with a free AI automation audit or see the full done-for-you automation services.

Meta Llama FAQ

How much does it cost to run Llama models?

It depends entirely on the host, since Meta publishes weights and others sell the inference. As of August 2026, Llama 3.1 8B is $0.05 in and $0.08 out per 1M on Groq and $0.02 and $0.04 on OpenRouter. Llama 3.3 70B is $0.10 and $0.32 on OpenRouter, $0.59 and $0.79 on Groq, and $1.04 flat on Together AI. Llama 4 Scout is $0.10 and $0.30 with a 1.31M context. We found no Llama 5 listed at any host we checked.

Why is the same Llama model ten times cheaper on one host than another?

Because you are buying different things. Groq documents 280 tokens per second on Llama 3.3 70B and charges $0.59 per 1M input; OpenRouter routes the same weights for $0.10 without that throughput guarantee. If a person is waiting on the output, the premium buys something real. If the job runs overnight, it does not. Decide whether you are buying tokens or speed before you pick a host, and re-check periodically because the spread moves.

What would a real Llama automation cost per month?

A 1,500-run inbox agent at about 3,000 input and 400 output tokens per run costs roughly $0.27 a month on Llama 3.1 8B via Groq, or about $0.11 via OpenRouter. A 4,000-document extraction job at 6,000 input and 300 output tokens each is about $1.30 and $0.53 respectively, or about $2.78 on Llama 3.3 70B via OpenRouter. Those are the lowest running costs in this directory, and they come with the most integration work.

Is there a Llama 5?

Not that we could verify. In August 2026 we found no Llama 5 on Groq's model list, on Together AI's serverless chat pricing, or in OpenRouter's Meta Llama catalogue, where Llama 4 Maverick and Scout are the newest entries. We are not going to quote pricing for a model we cannot find listed anywhere. If you need current-generation capability with verifiable hosted pricing, plan on Llama 4.

Is Llama actually open source?

Not in the strict sense. The Llama Community Licence is a custom licence, not an OSI-approved open-source licence, and it carries conditions including a very large monthly-user threshold and an acceptable-use policy. Weights are downloadable and commercial use and fine-tuning are permitted, so for an ordinary small business it rarely bites, but read it if you are embedding Llama in a product you sell.

Does Meta run its own Llama API?

Not one we could find or price. Meta's first-party paid endpoint appears to have been wound down, and llama.com now redirects to Meta's general developer site, which did not give us readable answers on generation, API availability or licence when we fetched it in August 2026. Treat that as our reading rather than a confirmed fact, but plan on choosing a host such as Groq, Together AI, Fireworks, OpenRouter or AWS Bedrock rather than expecting a Meta vendor relationship.

Can I get prompt caching or a batch discount on Llama?

Only if the host you chose offers them, and they are not model features. There is no equivalent of OpenAI's automatic caching or Anthropic's Batch API that follows the weights from host to host. This matters when comparing: a mainstream API with 90 percent caching and a 50 percent batch discount can close much of the apparent gap, so compare against its discounted rate rather than its list rate.

What hardware do I need to self-host Llama?

As a rough guide, an 8B-class model needs about 16GB of VRAM at full precision or 6 to 8GB quantised to four bits, which is a single consumer card. A 70B-class model needs about 140GB at full precision or roughly 40GB quantised. These are community rules of thumb rather than official Meta figures, so use them for planning and test before you buy.

Why does tool calling behave differently on different Llama hosts?

Because each host implements its own function-calling parser and schema handling on top of the same weights. An agent that works reliably on one host can misbehave on another. The mitigation is to use grammar-constrained or strict JSON output where available, validate every tool call against your schema before executing it, and include the host in your evaluation runs rather than only the model.

Should a small business use Llama or a mainstream API?

Use Llama when volume is high enough that per-token cost dominates, when latency matters enough that a fast specialist host wins, or when data cannot leave your network. Otherwise a mainstream API is usually the better deal, because one vendor, one bill and documented behaviour save more in build and maintenance time than cheap tokens save in spend.