Llama API pricing: every model, every rate
Llama is open-weight, so the same model is hosted by many providers at wildly different rates. Below is what each variant bills at here.
llama-4-maverick. Rates checked 2026-11-01.Full rate card
26 billable models in total, 11 of them text models. On those text models: 0 of 11 carry a rate verified against Meta's own published pricing, the remaining 11 are gateway rates only. USD per 1M tokens, checked 2026-11-01.
| Model | Input | Output | Out/In | Official (in / out) |
|---|---|---|---|---|
Llama-3.1-405B | $3 | $6 | 2.0× | not verified |
Meta-Llama-3.1-405B-Instruct | $3 | $6 | 2.0× | not verified |
llama-3.1-405b-instruct | $3 | $6 | 2.0× | not verified |
llama-3.2-90b-vision-instruct | $3 | $9 | 3.0× | not verified |
llama-3-70b | $2 | $4 | 2.0× | not verified |
llama-3.1-70b | $2 | $4 | 2.0× | not verified |
llama-3.1-70b-instruct | $2 | $2 | 1.0× | not verified |
meta-llama/llama-3.1-70b-instruct | $2 | $2 | 1.0× | not verified |
meta-llama/llama-4-maverick | $1.25 | $5 | 4.0× | not verified |
meta-llama/llama-4-scout | $1.25 | $5 | 4.0× | not verified |
llama-2-13b | $1 | $1 | 1.0× | not verified |
llama-2-70b | $1 | $1 | 1.0× | not verified |
llama-2-7b | $1 | $1 | 1.0× | not verified |
llama-3-8b | $1 | $1 | 1.0× | not verified |
llama-3.1-8b | $1 | $1 | 1.0× | not verified |
llama-3.2-11b-vision-instruct | $1 | $1 | 1.0× | not verified |
llama-3-sonar-large-32k-chat | $0.75 | $0.75 | 1.0× | not verified |
llama-3-sonar-small-32k-chat | $0.75 | $0.75 | 1.0× | not verified |
llama-3.2-3b-instruct | $0.5 | $0.25 | 0.5× | not verified |
llama-3.3-70b-instruct | $0.36 | $0.36 | 1.0× | not verified |
llama-3.1-70b-instruct-turbo | $0.25 | $1 | 4.0× | not verified |
llama-3.2-1b-instruct | $0.25 | $0.0625 | 0.2× | not verified |
llama-3.2-90b-vision | $0.17 | $0.1938 | 1.1× | not verified |
llama-3.1-8b-instruct | $0.125 | $0.5 | 4.0× | not verified |
llama-4-maverick | $0.07 | $0.35 | 5.0× | not verified |
llama-3.3-70b | $0.05 | $0.148 | 3.0× | not verified |
On Meta the Out/In column runs 2.0× to 5.0×, median 4.0×, it outweighs the input price as soon as output dominates your bill.
Across 11 priced text models the blended rate spans $0.198 (llama-3.3-70b) to $9 (Meta-Llama-3.1-405B-Instruct) per 1M tokens (a 45× spread). At 45× between its own cheapest and dearest, picking the wrong tier inside Meta costs more than picking the wrong vendor.
How these rates were verified
Meta publishes no USD rate card we could confirm against, so none of the 26 prices above is cross-checked with vendor documentation — every figure is a gateway rate read from the live rate config on 2026-11-01. That is the reason no discount percentage appears anywhere on this page: without a published list to compare against, a "saving" would be a number we made up. Read these as an observed snapshot of what the endpoint charges today, and treat the reason to route Meta through one endpoint as access and consolidated billing rather than a discount.
What Meta costs at three usage levels
Priced on llama-4-maverick at $0.07 in / $0.35 out per 1M tokens, with a 5:1 input-to-output ratio, roughly what an
interactive workload produces.
| Usage level | Tokens per month (in / out) | Monthly cost llama-4-maverick |
|---|---|---|
| Light pilot or side project | 5M / 1M | $0.70 |
| Working one product in production | 50M / 10M | $7.00 |
| Heavy high-volume pipeline | 500M / 100M | $70.00 |
This is the model we cover in depth for this vendor. Swap in your own token split in the calculator.
Same budget, other vendors
3 models from 3 other vendors priced within
35% of llama-4-maverick on a blended rate (the cheapest, the dearest and one
in between). This is the horizontal check a Meta rate page on its own cannot give you.
| Model | Vendor | In / out per 1M | vs this |
|---|---|---|---|
Claude-haiku-5-5 | Anthropic | $0.05 / $0.25 | −28.6% |
grok-3-mini | xAI | $0.15 / $0.2505 | −4.6% |
qwen3-30b-a3b-instruct-2507 | Alibaba | $0.1 / $0.4 | +19.0% |
Blended comparison (input + output per 1M tokens). Price is one axis, none of these is a Meta model, so treat them as candidates to test side by side at this budget rather than drop-in substitutes.
Why the output rate decides the bill
Across the 6 priced Meta text models, the output rate runs from 2.0× to 5.0× the input rate, with a median of 4.0×. On Meta that multiple lands on the smaller half of the token count and the larger half of the bill, because output is what a chat or agent turn actually produces.
Worked out: at a 4.0× multiple, a job whose token count is 80% input still spends 50% of its cost on the output it generates. Setting Meta beside any other vendor on the input rate alone hides that.
What Meta is good at
Llama 405B is the usual open-weight choice when you need on-prem-equivalent capability without a frontier price tag.
Detailed pages for Meta models
The models we cover in depth, each with worked cost at three usage levels and a comparison against similarly priced alternatives.
- Llama 4 Maverick — $0.07 / $0.35 per 1M tokens
Routing these rates into your tools
Every model above is reachable from the same OpenAI-compatible endpoint, so the rates on this page apply wherever you already work. These are the tools we have profiled end to end — Meta models included:
| Tool | What to set |
|---|---|
| LangChain | one base_url kwarg — then check whether token usage still comes back |
| LlamaIndex | OpenAILike for gateway IDs, and a second object for the embedding model |
| AutoGen | declare function_calling or tool use fails silently |
| OpenAI Agents SDK | tracing exports to OpenAI by default; the default transport is Responses |
| Codex CLI | 0.122+ rejects wire_api = "chat" at config load |
| Gemini CLI | a model without tool calls replies in prose and never touches the repo |
| CrewAI | custom_openai=True, and cost is agents × turns |
| DSPy | the bill is in compile(), not in inference |
| Cursor | override the base URL and bill your own key |
| LiteLLM model names | the slash is a routing instruction — openai/, or nothing is sent |
| streaming calls | stream_options, or every token count comes back null |
| an empty model picker | read /v1/models and type the id by hand |
| long sessions | two caps, not one — the window and the maximum output |
| every tool we have profiled | the full matrix of what to set and where |
Every one of these takes the same two values (a base URL and a key) and none of them require changing your code beyond where the client is constructed. Pointed at Meta, that is the whole change. The differences that matter are in each tool's defaults, which is what each guide covers.
Other vendors
Every rate card here is built the same way, from the same price snapshot (2026-11-01), so the numbers are comparable across vendors:
Anthropic · OpenAI · Google · DeepSeek · Alibaba · xAI · Zhipu · Moonshot · MiniMax · ByteDance
FAQ
How much does Meta charge per million tokens?
Across the 11 priced text models in this catalogue, the blended rate (input plus output per 1M tokens) runs from $0.198 on llama-3.3-70b to $9 on Meta-Llama-3.1-405B-Instruct, as of 2026-11-01. That 45× spread is what matters: inside Meta alone, picking the wrong tier moves the bill further than switching to another vendor would.
What is the cheapest Meta model for high-volume work?
llama-3.3-70b is the lowest blended rate here at $0.05 input and $0.148 output per 1M tokens. On a Heavy workload (500M input and 100M output tokens a month) that bills about $39.80. At the other end, the same workload on Meta-Llama-3.1-405B-Instruct costs about $2,100.00.
Is Meta cheaper through a gateway than buying direct?
We do not claim a discount on Meta. None of its 11 text models has a list price published in a stable machine-readable form, so all 11 figures on this page are gateway rates with nothing official to compare them against. Check the Meta number in your own dashboard before committing to a budget.
How does output pricing change a Meta bill?
On Meta output is billed at 2.0× to 5.0× the input rate, median 4.0×, so the input rate alone never predicts the bill. At 4.0× even a workload that is 80% input by token count lands most of its cost on the output column. Price your own split against llama-4-maverick's $0.07 input rate in the calculator rather than extrapolating from that headline.
What is the Meta API price per 1M tokens?
Meta quotes every model separately for input and output. llama-4-maverick, the one closest to this vendor's median rate, is $0.07 per 1M input and $0.35 per 1M output as of 2026-11-01. The lowest tier in this catalogue is llama-3.3-70b, at $0.05 in and $0.148 out.
How much does Meta cost for a coding agent?
A heavy agent month — 500M input and 100M output tokens, the shape a coding assistant really produces because every turn resends the whole context (bills about $70).00 on llama-4-maverick. Output decides most of that figure: on Meta the median model bills output at 4.0× its input rate, so extrapolating from the headline input rate understates the bill by close to that factor.
Which provider has the lowest-cost Meta API?
In this catalogue llama-3.3-70b carries the lowest blended rate at $0.05 per 1M input and $0.148 per 1M output. Because these models are open-weight, the identical weights are hosted by many providers at different margins — the spread between hosts is mostly infrastructure cost rather than any difference in the model, so it is worth checking more than one before committing. Rates checked 2026-11-01.
Are the Llama models here the cheapest way to run open-weight work?
Among open-weight families in this catalogue they are the lowest-priced, with the 70B-class tiers in the sub-$1 blended band. The gap to the cheapest closed models is small, so the decision usually comes down to whether you need the weights themselves.
See all 38 models → or use the cost calculator with your own token split.