BroadRouter

Llama 3.3 70B API pricing

Updated

Among the 14 providers we checked (prices read 2026-11-14 to 2026-11-15), the lowest blended standard price for Llama 3.3 70B is DeepInfra at USD 0.10 per 1M input tokens and USD 0.32 per 1M output tokens. The median across 14 standard prices is USD 0.625 / USD 0.72. Meta sells no API of its own for it.

Meta · slug llama-3.3-70b-instruct · up to 131,072 tokens of context · 14 providers with a standard price. Every price is read from the provider's own pricing page and linked in its row, checked 2026-11-15. No aggregator feed. Download all prices (every provider and model, with sources, CSV or JSON). How it compares with the market: the LLM API price index. Archived list prices since 2025-07-09 (4 archived price points, 0 list-price changes): LLM API price history.

Does Meta sell Llama 3.3 70B directly?

No. Meta does not sell Llama models through an API of its own: llama.com now redirects to Meta's developer site, whose Model API lists only Meta's Muse models. Llama is available only from third-party hosts. Checked 2026-11-14.

Llama 3.3 70B model facts

Facts from Meta's model card for meta-llama/Llama-3.3-70B-Instruct and the Hugging Face model API, read 2026-11-16. BroadNet did not verify them, and hosts may serve the model at a different precision. BroadRouter is BroadNet's site.

Llama 3.3 70B: facts as stated by the model card and the Hugging Face model API, with sources
FactAs statedSource
Open weightsYes, with gated access. meta-llama/Llama-3.3-70B-Instruct is public, but Hugging Face asks you to request access and accept the terms before you download (gated: manual).Hugging Face model API, checked 2026-11-16
Quote
“"private": false, "gated": "manual", 30 .safetensors files listed”
LicenceLlama 3.3 Community License Agreement (custom commercial license)Meta model card, checked 2026-11-16
Quote
“A custom commercial license, the Llama 3.3 Community License Agreement”
Total parameters70BMeta model card, checked 2026-11-16
Quote
“Llama 3.3 (text only) | A new mix of publicly available online data. | 70B”
Context length128kMeta model card, checked 2026-11-16
Quote
“Multilingual Text and code | 128k”
Input and outputMultilingual text in; multilingual text and code outMeta model card, checked 2026-11-16
Quote
“Input modalities … Output modalities … Multilingual Text … Multilingual Text and code”
Tool callingYesMeta model card, checked 2026-11-16
Quote
“LLaMA-3.3 supports multiple tool use formats.”
Release date2024-12-06Meta model card, checked 2026-11-16
Quote
“70B Instruct: December 6, 2024”
First published on Hugging Face (not necessarily the release date)2024-11-26Hugging Face model API, checked 2026-11-16
Quote
“"createdAt": "2024-11-26T16:08:47.000Z"”

Llama 3.3 70B price per provider

USD per 1M tokens. Standard prices are ranked on a blend of 3 input tokens to 1 output token, cheapest first; the dearest standard price is 7.9× the cheapest. Free rows, the maker's serving version and other service tiers follow, tagged and unranked.

Llama 3.3 70B API prices by provider
#ProviderInputCached inputOutputBlended 3:1ContextPrecisionSource
1DeepInfra cheapest
meta-llama/Llama-3.3-70B-Instruct-Turbo
Notes and source quote

DeepInfra Turbo variant (quantized).

“$0.10 in $0.32 out / 1M tokens”
USD 0.10not publishedUSD 0.32USD 0.155131,072FP8DeepInfra page
checked 2026-11-14
re-checked 2026-11-16
2Novita AI
meta-llama/llama-3.3-70b-instruct
Notes and source quote
“Llama 3.3 70B Instruct 12K $0.135 /Mt $0.4 /Mt”
USD 0.135not publishedUSD 0.40USD 0.201312,288not statedNovita AI page
checked 2026-11-14
re-checked 2026-11-16
3AkashML
meta-llama/Llama-3.3-70B-Instruct
Notes and source quote
“Price (per 1M Tokens) Input Output Cache Read Llama 3.3 70B $0.20 $0.52 $0.10”
USD 0.20USD 0.10USD 0.52USD 0.28131,072not statedAkashML page
checked 2026-11-15
re-checked 2026-11-16
4Parasail
Notes and source quote
“Llama 3.3 70B (FP8) $0.22 $0.50 $0.11”
USD 0.22USD 0.11USD 0.50USD 0.29not statedFP8Parasail page
checked 2026-11-14
re-checked 2026-11-16
5GMI Cloud
meta-llama/Llama-3.3-70B-Instruct
Notes and source quote

No cache-read price published.

“Price (per 1M tokens) Input $0.25 Output $0.75”
USD 0.25not publishedUSD 0.75USD 0.375131,072BF16GMI Cloud page
checked 2026-11-15
re-checked 2026-11-16
6Featherless
meta-llama/Llama-3.3-70B-Instruct
Notes and source quote

Developer plan (per-token billing, from 50 USD a month in credits).

“PRICING: INPUT $0.65 / CACHED $0.52 / OUTPUT $0.75”
USD 0.65USD 0.52USD 0.75USD 0.67532,768FP8Featherless page
checked 2026-11-15
re-checked 2026-11-16
7CoreWeave Forge
meta-llama/Llama-3.3-70B-Instruct
Notes and source quote
“Meta Llama 3.3 70B $0.71 $0.71 -”
USD 0.71not publishedUSD 0.71USD 0.71128,000not statedCoreWeave Forge page
checked 2026-11-15
re-checked 2026-11-16
8Microsoft Azure
Notes and source quote

Foundry Models serverless (pay-as-you-go), Global deployment.

“Llama 3.3 70B Global $0.71 $0.71”
USD 0.71not publishedUSD 0.71USD 0.71not statednot statedMicrosoft Azure page
checked 2026-11-14
re-checked 2026-11-16
9Amazon Bedrock
Notes and source quote

Shown for region US East (N. Virginia) on the pricing page (values filled from AWS's published price list for us-east-1).

“Llama 3.3 Instruct (70B) $0.72 $0.72 $0.36 $0.36”
USD 0.72not publishedUSD 0.72USD 0.72not statednot statedAmazon Bedrock page
checked 2026-11-14
re-checked 2026-11-16
10Google Vertex AI
llama-3.3-70b
Notes and source quote
“Llama 3.3 70B Input $0.72 … Output $0.72”
USD 0.72not publishedUSD 0.72USD 0.72not statednot statedGoogle Vertex AI page
checked 2026-11-14
re-checked 2026-11-16
11SambaNova
Meta-Llama-3.3-70B-Instruct
Notes and source quote
“Meta-Llama-3.3-70B-Instruct N/A 0.60 USD 1.20 USD”
USD 0.60not publishedUSD 1.20USD 0.75not statednot statedSambaNova page
checked 2026-11-14
re-checked 2026-11-16
12IBM watsonx.ai
llama-3-3-70b-instruct
Notes and source quote

Single published per-1M rate; IBM bills Resource Units "equivalent to 1000 tokens (including both input and output tokens)", so input and output cost the same.

“llama-3-3-70b-instruct Meta USD 0.7526”
USD 0.7526not publishedUSD 0.7526USD 0.7526not statednot statedIBM watsonx.ai page
checked 2026-11-15
re-checked 2026-11-16
13Cloudflare Workers AI
@cf/meta/llama-3.3-70b-instruct-fp8-fast
Notes and source quote

Cloudflare publishes this USD per-token price alongside the neuron price.

“@cf/meta/llama-3.3-70b-instruct-fp8-fast $0.293 per M input tokens $2.253 per M output tokens”
USD 0.293not publishedUSD 2.253USD 0.783not statedFP8Cloudflare Workers AI page
checked 2026-11-14
re-checked 2026-11-16
14Venice
llama-3.3-70b
Notes and source quote
“| Llama 3.3 70B | `llama-3.3-70b` | $0.70 | $2.80 | — | — | Private |”
USD 0.70not publishedUSD 2.80USD 1.225128,000FP8Venice page
checked 2026-11-15
re-checked 2026-11-16
—Amazon Bedrock batch
Notes and source quote

Batch columns. Shown for region US East (N. Virginia) on the pricing page (values filled from AWS's published price list for us-east-1).

“Llama 3.3 Instruct (70B) $0.72 $0.72 $0.36 $0.36”
USD 0.36not publishedUSD 0.36USD 0.36not statednot statedAmazon Bedrock page
checked 2026-11-14
re-checked 2026-11-16
—Google Vertex AI batch
llama-3.3-70b
Notes and source quote
“Llama 3.3 70B Batch Input $0.36 … Batch Output $0.36”
USD 0.36not publishedUSD 0.36USD 0.36not statednot statedGoogle Vertex AI page
checked 2026-11-14
re-checked 2026-11-16
—DeepInfra flex / spot
meta-llama/Llama-3.3-70B-Instruct-Turbo
Notes and source quote

Flex service tier (0.8× standard). DeepInfra Turbo variant (quantized).

“Flex (0.8×) $0.08 $0.256”
USD 0.08not publishedUSD 0.256USD 0.124131,072FP8DeepInfra page
checked 2026-11-14
re-checked 2026-11-16

Report a wrong price or number (email to info@broadnet.ai, with this page's link filled in).

What does Llama 3.3 70B cost for a real workload?

Monthly cost over 30 days at each standard price, using the presets of the LLM API cost calculator. Providers without a cached price bill cached input at the input price.

  • Support chatbot: 2,000 requests a day, 1,500 input tokens (60% cached), 300 output.
  • RAG over documents: 1,000 requests a day, 8,000 input tokens (30% cached), 500 output.
  • Coding agent (one developer): 1,500 requests a day, 40,000 input tokens (85% cached), 2,000 output incl. reasoning.
Monthly cost of three workloads on Llama 3.3 70B, by provider
ProviderSupport chatbotRAG over documentsCoding agent (one developer)
DeepInfraUSD 14.76USD 28.80USD 208.80
Novita AIUSD 19.35USD 38.40exceeds context
AkashMLUSD 21.96USD 48.60USD 253.80
ParasailUSD 22.86USD 52.38USD 272.70
GMI CloudUSD 36.00USD 71.25USD 517.50
FeatherlessUSD 64.98USD 157.89exceeds context
CoreWeave ForgeUSD 76.68USD 181.05USD 1,342
Microsoft AzureUSD 76.68USD 181.05USD 1,342
Amazon BedrockUSD 77.76USD 183.60USD 1,361
Google Vertex AIUSD 77.76USD 183.60USD 1,361
SambaNovaUSD 75.60USD 162.00USD 1,188
IBM watsonx.aiUSD 81.28USD 191.91USD 1,422
Cloudflare Workers AIUSD 66.92USD 104.11USD 730.17
VeniceUSD 113.40USD 210.00USD 1,512

For the support chatbot, the cheapest price (DeepInfra, USD 14.76 a month) saves USD 56.50 against the median (USD 71.26). For the coding agent the gap is USD 904.50 a month (81% lower).

Which Llama 3.3 70B provider is cheapest for your traffic?

DeepInfra has the lowest input price (USD 0.10) and the lowest output price (USD 0.32), so it is the cheapest standard price at every mix of input and output tokens, before cache discounts.

Llama 3.3 70B: price per 1M tokens at 3 input to 1 output, by share of input served from cache
ProviderNo cache hits50% cached90% cached
DeepInfra (no cache price)USD 0.155 cheapestUSD 0.155 cheapestUSD 0.155 cheapest
Novita AI (no cache price)USD 0.2013USD 0.2013USD 0.2013
AkashMLUSD 0.28USD 0.2425USD 0.2125
ParasailUSD 0.29USD 0.2487USD 0.2157
GMI Cloud (no cache price)USD 0.375USD 0.375USD 0.375
FeatherlessUSD 0.675USD 0.6262USD 0.5873

USD per 1M tokens, 3 input tokens to 1 output token, each provider's cheapest standard row; 8 dearer providers are in the price table above. A provider without a cached price bills cached input at the input price. Any charge for writing to the cache, which some providers add, is not included.

3 of 14 providers publish a cached-input price below the input price: AkashML (50% off), Parasail (50% off), Featherless (20% off). DeepInfra, Novita AI, GMI Cloud, CoreWeave Forge, Microsoft Azure and 6 more publish none, so cached tokens there bill at the input price.

Context windows differ: 131,072 tokens at DeepInfra, AkashML and GMI Cloud; 128,000 tokens at CoreWeave Forge and Venice; 32,768 tokens at Featherless; 12,288 tokens at Novita AI. Parasail, Microsoft Azure, Amazon Bedrock, Google Vertex AI, SambaNova and 2 more do not state one. The cheapest row is not always the one that fits a long prompt.

What does Llama 3.3 70B cost in other languages?

The same text needs a different number of tokens in each language. With the Llama 3.3 70B tokenizer, sending 1M characters costs this much at the cheapest standard input price we found for Llama 3.3 70B:

ContentInput per 1M charactersCheapest provider
EnglishUSD 0.022DeepInfra
Arabic MSAUSD 0.039DeepInfra
HindiUSD 0.052DeepInfra
ChineseUSD 0.080DeepInfra
CodeUSD 0.025DeepInfra

From BroadRouter's real cost by language dataset, which covers 15 languages and content types for every model we could match to a measured tokenizer.

Published benchmark results

BroadNet did not run these benchmarks; results are as published by the Meta model card and Epoch AI and may use different settings. BroadRouter is BroadNet's site.

Published benchmark results for Llama 3.3 70B, with sources
BenchmarkResultSetting or noteSource
MMLU Pro (CoT)68.95-shot, chain of thoughtMeta model card, checked 2026-11-16
IFEval92.1not statedMeta model card, checked 2026-11-16
GPQA Diamond (CoT)50.50-shot, chain of thoughtMeta model card, checked 2026-11-16
HumanEval88.40-shot, pass@1Meta model card, checked 2026-11-16
GPQA Diamond0.474Setting as listed by Epoch (no suffix); Epoch run Llama-3.3-70B-Instruct, mean score on a 0 to 1 scaleEpoch AI, checked 2026-11-16

Epoch AI rows: Epoch AI, Capabilities & benchmarking, licensed CC BY 4.0, shown as the mean score on a 0 to 1 scale, rounded to three decimals (hover a value for the exact figure in Epoch's data file).

Estimate your own bill

The Meta models whose cheapest standard price is closest to Llama 3.3 70B's (USD 0.155 blended):

Meta models priced closest to Llama 3.3 70B
ModelCheapest inputCheapest outputMax contextProviders
Llama 3.2 11B Vision InstructUSD 0.16USD 0.16—3
Llama 4 Scout 17B 16E InstructUSD 0.10USD 0.301,048,5766
Llama 3.2 3B InstructUSD 0.0509USD 0.335128,0004
Llama 3 8B InstructUSD 0.0925USD 0.0958,1923
Llama 3.2 1B InstructUSD 0.027USD 0.20132,7683
Llama 4 Maverick 17B 128E Instruct FP8USD 0.25USD 0.801,048,5767

Priced closest from other makers: NVIDIA Nemotron 3 Super 120B A12B (NVIDIA, USD 0.1563) · Qwen3 235B A22B Instruct 2507 FP8 (Qwen, USD 0.1531) · gemma-4-31B-it-turbo (Google, USD 0.1525).

Compare Llama 3.3 70B with

Also compared: vs Llama 4 Maverick 17B 128E Instruct FP8 · vs OpenAI GPT OSS 120B · vs Qwen2.5-72B-Instruct · vs Gemma 4 26B A4B · vs Llama 3.1 Instruct (70B).

Every Llama 3.3 70B price comparison · All model-vs-model price comparisons

All 428 models: API prices compared · Cheapest LLM APIs right now · LLM API price index (open data) · Download all prices (CSV, JSON) · Price change log

Questions

How much does the Llama 3.3 70B API cost per token?

From USD 0.10 per 1M input tokens (USD 0.0000001 per token) and USD 0.32 per 1M output tokens (USD 0.00000032 per token) at the lowest standard prices we found. The median of 14 standard prices is USD 0.625 input and USD 0.72 output. Each price was read from the provider's own pricing page; the source is linked in its row.

Which provider is cheapest for Llama 3.3 70B?

Among the 14 providers we checked (prices read 2026-11-14 to 2026-11-15), the lowest is DeepInfra, ranked on a blend of 3 input tokens to 1 output token (USD 0.155 per 1M). It also has the lowest input and the lowest output price. Batch, flex, priority and fast tiers are shown in the table but not ranked.

Does Meta sell Llama 3.3 70B on its own API?

Meta does not sell Llama models through an API of its own: llama.com now redirects to Meta's developer site, whose Model API lists only Meta's Muse models. Llama is available only from third-party hosts. The lowest-priced host we checked is DeepInfra at USD 0.10 / USD 0.32 per 1M tokens.

Does Llama 3.3 70B have cached input pricing?

3 of 14 standard prices include one. The lowest is USD 0.10 per 1M at AkashML; the median is USD 0.11. Providers without one bill cached input at the input price. What cached input means.

What is the Llama 3.3 70B context window?

Up to 131,072 tokens. 4 standard prices come with less, down to 12,288 tokens (Novita AI, Featherless, CoreWeave Forge and Venice). Check the context column before you pick the cheapest row for long prompts. Context window calculator.

How much does a coding agent cost per month on Llama 3.3 70B?

For 1,500 requests a day, 40,000 input tokens (85% cached), 2,000 output incl. reasoning, over 30 days: USD 208.80 a month at DeepInfra, USD 1,113 at the median price and USD 1,512 at Venice. Try your own numbers.

Is there a batch, flex or priority price for Llama 3.3 70B?

Yes: batch from Amazon Bedrock at USD 0.36 / USD 0.36; flex / spot from DeepInfra at USD 0.08 / USD 0.256 per 1M input / output tokens. These tiers trade speed or availability for price (batch runs within hours, flex and spot can be slow or refused at capacity, priority and fast cost more for speed or reliability), so they are shown but not ranked.

Is Llama 3.3 70B free to use through an API?

No provider in our data lists a free price for Llama 3.3 70B as of 2026-11-15. Free tiers and credits by provider (checked 2026-11-14) lists general offers that do not name a model; all free LLM API models are on one page.

Is Llama 3.3 70B open source, with open weights, and what is its licence?

Meta publishes the weights on Hugging Face as meta-llama/Llama-3.3-70B-Instruct, with gated access: Hugging Face asks you to request access and accept the terms before you download. The licence, as the Meta model card states it: Llama 3.3 Community License Agreement (custom commercial license). Open weights means you can download and run the model yourself; what you may do with it, including commercial use, is set by that licence, so read it in full (checked 2026-11-16). These facts come from Meta's card and the Hugging Face model API; BroadNet did not verify them, and hosts may serve the model at a different precision.

How many parameters does Llama 3.3 70B have?

Total parameters, as the Meta model card states them: 70B. (checked 2026-11-16; BroadNet did not verify these figures.)

Does Llama 3.3 70B support tool calling?

Yes, according to the Meta model card: “LLaMA-3.3 supports multiple tool use formats.” (checked 2026-11-16). Whether the tools parameter works through an API also depends on the provider serving it, so check that provider's docs. BroadNet did not test it.

How do I call Llama 3.3 70B with the OpenAI SDK?

BroadNet does not serve Llama 3.3 70B. Use the OpenAI-compatible base URL of a provider that lists it: create the client with OpenAI(base_url=..., api_key=...) and pass that provider's model ID. Documented base URLs for providers on this page, in alphabetical order: DeepInfra (https://api.deepinfra.com/v1/openai, model meta-llama/Llama-3.3-70B-Instruct-Turbo). Each provider page quotes the provider's docs with a check date.

How does Llama 3.3 70B score on benchmarks?

As published, the Meta model card reports 68.9 on MMLU Pro (CoT) and 92.1 on IFEval; Epoch AI measured 0.474 on GPQA Diamond in its own run, as a mean score on a 0 to 1 scale (checked 2026-11-16). BroadNet did not run these benchmarks, and sources may use different settings, so the figures are not directly comparable. Every published result we found, with its setting and source.

Read as MarkdownAsk ChatGPTAsk ClaudeAsk Perplexity