Llama 4 Scout API Rate Limits & Pricing (July 2026)

MetaGenerally available

API model string

meta-llama/Llama-4-Scout-17B-16E-Instruct (host-dependent)

Context window

10M tokens (native); 128K (Groq), 320K (DeepInfra), ~1M (Fireworks); full 10M only via self-hosting

Max output not published (host-capped)

Knowledge cutoff: not published
Released: Apr 2026
Modalities: text, image

Last verified July 10, 2026

Rate limits by tier

Llama 4 Scout has no first-party hosted API from Meta (waitlist only). Rate limits are entirely set by the inference host. Groq offers a free tier at ~30 RPM; paid hosts do not publicly publish RPM limits — always check your org's console. The 10M native context is only available via self-hosting; hosted caps range from 128K (Groq) to ~1M (Fireworks).

Tier	Requirements	RPM	TPM	RPD	Concurrent	Notes
Meta first-party	Waitlist only — not an active commercial endpoint	not published	not published	not published	not published	Meta's own Llama API requires waitlist access; Llama 4 Scout is not an active first-party commercial endpoint as of July 2026
Groq (third-party)	Free tier available — no credit card required	~30 (free tier); higher on paid plans	not published	not published	not published	$0.11 in / $0.34 out per MTok (verified July 10, 2026). Fastest inference host at ~446 t/s (Artificial Analysis, July 2026). Context capped at 128K — use DeepInfra or Fireworks for longer contexts.
DeepInfra (third-party)	Pay-as-you-go — credit balance required	not published	not published	not published	not published	$0.08 in / $0.30 out per MTok — cheapest available host (verified July 10, 2026). 320K context exposed. Best option for cost-sensitive workloads.
Fireworks AI (third-party)	Pay-as-you-go	not published	not published	not published	not published	Context ~1M exposed (closest to native of any listed host); higher throughput than Groq for long-context workloads. Pricing: verify at fireworks.ai before committing.
CompactifAI (third-party)	Pay-as-you-go — verify availability and SLA before production use	not published	not published	not published	not published	$0.10 in / $0.14 out per MTok — cheapest blended price among listed hosts (verify availability and SLA before production use)
Together AI (third-party)	Pay-as-you-go	not published	not published	not published	not published	$0.15 in / $0.60 out per MTok (verified July 10, 2026). Pricier than DeepInfra at same token volume.

Swipe the table sideways to see every limit column.

1.All rate limits are set by the inference host, not Meta. HTTP 429 on all OpenAI-compatible hosts returns a Retry-After header.
2.The 10M native context window is only available via self-hosting. Hosted caps: Groq 128K, DeepInfra 320K, Fireworks ~1M. Check the host's console before building long-context workflows.
3.Llama 4 Community License: platforms with >700M MAU require a separate agreement with Meta. Inference speed on Groq (~446 t/s) is fastest among surveyed hosts as of July 10, 2026 (Artificial Analysis).

Limits verified against the Meta docs, July 10, 2026.

Token pricing

What you pay per million tokens (USD). Input and output are billed separately.

Input

$0.08

per 1M tokens

Output

$0.30

per 1M tokens

Prices shown are for DeepInfra ($0.08/$0.30 per MTok) — the cheapest verified host as of July 10, 2026.
Groq: $0.11 in / $0.34 out at ~446 t/s (verified July 10, 2026). Together AI: $0.15 in / $0.60 out (verified July 10, 2026). CompactifAI: $0.10 in / $0.14 out (cheapest blended — verify availability before production).
No standard cross-host prompt caching for Llama 4 Scout. The 70%-cached scenario is not applicable. Cache system prompts at the application layer.
Amazon Bedrock offers batch inference at ~50% off for eligible throughput. Llama 4 Community License applies — platforms with >700M MAU need a separate Meta agreement.

Side-project (5M in / 1M out)

$0.70

per month

Assumptions

DeepInfra $0.08 in / $0.30 out — cheapest verified host; no standard caching

5M × $0.08 + 1M × $0.30 = $0.40 + $0.30 = $0.70. Groq alternative: 5M × $0.11 + 1M × $0.34 = $0.55 + $0.34 = $0.89.

Mid-scale app (60M in / 12M out)

$8.40

per month

Assumptions

DeepInfra $0.08/$0.30; no standard caching — 70%-cached scenario not applicable

60M × $0.08 + 12M × $0.30 = $4.80 + $3.60 = $8.40. No cache discount on Llama 4 hosts. Compare Llama 4 Maverick on DeepInfra: $9.00 + $7.20 = $16.20 — Scout is 48% cheaper.

High-volume (400M in / 80M out)

$56

per month

Assumptions

DeepInfra $0.08/$0.30; no standard caching

400M × $0.08 + 80M × $0.30 = $32 + $24 = $56. At CompactifAI $0.10/$0.14 (verify): $40 + $11.20 = $51.20 — even cheaper blended if SLA is acceptable.

Run your own numbers

Drag your real monthly token volumes and watch the bill update live — priced against rival models at the same usage.

Input tokens / month30M

1M500M

Output tokens / month5M

0.1M100M

Estimated meta-llama/Llama-4-Scout-17B-16E-Instruct (host-dependent) spend

$3.90/mo

Input: $2.40

Output: $1.50

30M in × $0.080 + 5M out × $0.300 = $3.90

Same volume, priced across models

Gemma 3Cheapest$1.60
meta-llama/Llama-4-Scout-17B-16E-Instruct (host-dependent)This model$3.90
Llama 4 Maverick$7.50

Rivals priced at their published input/output rates for the same monthly volumes. Prompt caching is model-specific, so it is applied to meta-llama/Llama-4-Scout-17B-16E-Instruct (host-dependent) only. Estimates for comparison; real bills vary with request shape and long-context surcharges.

meta-llama/Llama-4-Scout-17B-16E-Instruct (host-dependent) vs the alternatives

Llama 4 Scout compared to Llama 4 Maverick (heavier, pricier) and Gemma 3 (Google, smaller and cheaper at entry level) as of July 2026.

Aspect	meta-llama/Llama-4-Scout-17B-16E-Instruct (host-dependent)	Llama 4 Maverick	Gemma 3
Input price (cheapest host)	$0.08/MTok (DeepInfra)	$0.15/MTok (DeepInfra)	$0.04/MTok (Gemma 3 4B via Google)
Output price (cheapest host)	$0.30/MTok (DeepInfra)	$0.60/MTok (DeepInfra)	$0.08/MTok (Gemma 3 4B)
Native context window	10M	1M	131K (4B)
Hosted context (typical best)	128K–320K (Groq/DeepInfra); ~1M (Fireworks)	~1M (Fireworks)	131K
Reasoning quality	Lighter (16 experts, 109B total)	Stronger (128 experts, 400B total)	Strong for size
Multimodal	Text + image	Text + image	Text + image (4B+)
Inference speed	~446 t/s on Groq (fastest surveyed)	not published per host	not published
License	Llama 4 Community (>700M MAU needs Meta)	Same	Gemma Terms of Use

Swipe the table sideways to see every model.

Hitting a 429? The playbook

The exact errors you'll see

429 Too Many Requestsrate_limit_exceededRateLimitErrorThrottlingException

Why it happens & how to fix it

Groq free tier RPM exhausted (~30 RPM on Llama 4 Scout)

Upgrade to Groq paid tier, or switch to DeepInfra pay-as-you-go at $0.08/MTok input — typically lower effective rate limits for paying customers.

Context exceeds Groq's 128K cap, causing errors or silent truncation

Move long-context requests to DeepInfra (320K) or Fireworks (~1M). Do not assume the 10M native context is available on Groq.

Host TPM cap hit during a burst of concurrent requests

Implement exponential backoff with jitter; distribute requests across DeepInfra + Fireworks using round-robin routing with health-check failover.

CompactifAI availability or SLA issues causing 429 or timeouts

Fail over to DeepInfra ($0.08/MTok) or Groq. CompactifAI offers the cheapest blended price but verify SLA before relying on it in production.

Amazon Bedrock ThrottlingException — per-model account quota reached

Go to AWS Console → Service Quotas → Amazon Bedrock → Llama 4 Scout → request quota increase. Approval typically takes 1–3 business days.

Retry strategy

Use exponential backoff with jitter and honor the host's Retry-After header. All OpenAI-compatible hosts (Groq, DeepInfra, Fireworks, Together) return this header on 429. Formula: wait = min(2^attempt + random(0,1), 60). Max 5 retries. With OpenAI SDK: set max_retries=5 for automatic handling.

retry.ts

1import OpenAI from 'openai';
2
3// Choose your host's base URL:
4// DeepInfra (cheapest): https://api.deepinfra.com/v1/openai
5// Groq (fastest):       https://api.groq.com/openai/v1
6// Fireworks (~1M ctx):  https://api.fireworks.ai/inference/v1
7const client = new OpenAI({
8  baseURL: 'https://api.deepinfra.com/v1/openai',
9  apiKey: process.env.HOST_API_KEY,
10  maxRetries: 0, // handle retries manually for full control
11});
12
13async function callWithRetry(
14  messages: OpenAI.Chat.ChatCompletionMessageParam[],
15  maxRetries = 5
16): Promise<string> {
17  for (let attempt = 0; attempt < maxRetries; attempt++) {
18    try {
19      const response = await client.chat.completions.create({
20        model: 'meta-llama/Llama-4-Scout-17B-16E-Instruct',
21        messages,
22      });
23      return response.choices[0].message.content ?? '';
24    } catch (err: any) {
25      if (err?.status === 429) {
26        const retryAfter = parseInt(err?.headers?.['retry-after'] ?? '0', 10);
27        const jitter = Math.random();
28        const waitMs = retryAfter > 0
29          ? retryAfter * 1000
30          : Math.min(Math.pow(2, attempt) + jitter, 60) * 1000;
31        console.warn(`429 rate limit — attempt ${attempt + 1}/${maxRetries}, waiting ${Math.round(waitMs)}ms`);
32        await new Promise(r => setTimeout(r, waitMs));
33      } else {
34        throw err;
35      }
36    }
37  }
38  throw new Error('Max retries exceeded after repeated 429 responses');
39}

How to raise your limits

The ladder from the starter tier to enterprise — what each rung takes, and what it unlocks.

Free (Groq)

Immediate

Unlocks: ~30 RPM, 128K context, fastest inference (~446 t/s). Best for development and low-volume production.

Pay-as-you-go (DeepInfra)

Immediate on payment

Add credit balance at deepinfra.com

Unlocks: $0.08/MTok input — cheapest available. 320K context, higher RPM for paying customers.

Pay-as-you-go (Fireworks)

Immediate on payment

Add credit balance at fireworks.ai

Unlocks: ~1M context (closest to the native 10M cap for hosted access), competitive pricing (~$0.17/MTok input).

Amazon Bedrock

1–3 business days for quota increase

Enable Llama 4 Scout model access in Bedrock console → request throughput increase via AWS Service Quotas

Unlocks: Batch inference at ~50% off, AWS ecosystem integration, SLAs, enterprise compliance.

Self-host

Days to weeks

Download open weights under Llama 4 Community License; deploy on GPU cluster. Llama 4 Scout has 17B active parameters (16 experts from 109B total) — significantly lower GPU memory requirement than dense 70B or 405B models.

Unlocks: Full 10M context window, zero per-token cost, complete control over throughput. Requires separate Meta agreement if platform has >700M MAU.

Cut your token spend

Groq for latency-critical tasks, DeepInfra for cost-critical tasks

Groq: ~446 t/s (fastest); DeepInfra: $0.08/MTok (31% cheaper than Groq's $0.11)

Route streaming chat and interactive tasks to Groq. Route batch summarization, classification, and background tasks to DeepInfra. Both support the same model string with a base_url swap.

Use DeepInfra (320K) or Fireworks (~1M) for contexts above 128K

Avoids context truncation errors that occur silently on Groq's 128K cap

Check the requested context length before routing. If context > 100K tokens, route to DeepInfra. If context > 300K tokens, route to Fireworks.

Use Scout instead of Maverick for 80%+ of tasks

~47% cheaper on input ($0.08 vs $0.15) and 50% cheaper on output ($0.30 vs $0.60)

Scout handles classification, summarization, extraction, RAG retrieval, and conversational tasks as well as Maverick in most evaluations. Escalate to Maverick only for complex multi-step reasoning.

Amazon Bedrock batch for bulk non-real-time inference

~50% off on eligible throughput

Submit document processing, evals, and bulk content generation as Bedrock batch jobs rather than synchronous API calls. Results are delivered asynchronously to S3.

Cache system prompts at the application layer

Typically 20–40% reduction in repeated token spend

No native cross-host prompt caching exists for Llama 4 Scout. Store the system prompt in your backend and avoid re-sending it with every request. Use a consistent prefix to maximize cache reuse.

Consider CompactifAI for lowest blended cost (verify first)

$0.10 in / $0.14 out — cheapest blended price; could cut output costs by 53% vs DeepInfra

CompactifAI offers the lowest blended rate among listed hosts. Verify SLA, uptime, and model availability before moving production traffic. Use DeepInfra as the fallback.

Self-host for full 10M context

Zero per-token cost at scale; full 10M native context window

Llama 4 Scout's MoE architecture (17B active of 109B total) uses significantly less GPU memory than dense models of equivalent total size. Evaluate self-hosting at roughly $400–700/month of API spend, depending on GPU costs.

Frequently asked questions

What is Llama 4 Scout's API rate limit?

Llama 4 Scout has no first-party Meta API. Rate limits are set entirely by your inference host. Groq's free tier is approximately 30 RPM with a 128K context cap. DeepInfra, Fireworks, and Together AI do not publish RPM limits publicly — check your organization's console. All hosts return 429 with a Retry-After header when limits are exceeded.

Is Llama 4 Scout free to use?

Groq offers Llama 4 Scout on a free tier with approximately 30 RPM at $0.11/$0.34 per MTok. There is no zero-cost free inference tier. For production, DeepInfra at $0.08/$0.30 per MTok is the cheapest paid option (verified July 10, 2026).

How do I access Llama 4 Scout's full 10M context window?

The 10M native context is only available via self-hosting the open weights. Hosted providers cap the context window: Groq at 128K, DeepInfra at 320K, and Fireworks at approximately 1M. Fireworks is the best hosted option for long-context workloads — use it for requests over 320K tokens.

Llama 4 Scout vs Llama 4 Maverick: which should I choose?

Scout for most use cases — it is 47% cheaper on input ($0.08 vs $0.15/MTok) and 50% cheaper on output ($0.30 vs $0.60/MTok) on DeepInfra. Maverick has stronger reasoning capability (128 experts vs 16, 400B total vs 109B) and a 1M native context window. Route tasks to Scout by default and escalate to Maverick only for complex multi-step reasoning where quality differences are measurable.

How do I increase Llama 4 Scout rate limits?

Since Meta does not operate a first-party API, rate limits are managed on each inference host. To increase limits: upgrade from Groq's free tier to a paid plan; add credit balance on DeepInfra or Fireworks; request a quota increase on Amazon Bedrock via AWS Service Quotas (1–3 business days); or distribute load across multiple hosts. For help architecting a multi-host routing system with higher effective throughput, RapidDev offers free scoping calls at rapidevelopers.com/contact.

What is the cheapest way to run Llama 4 Scout in production?

DeepInfra at $0.08 in / $0.30 out per MTok is the cheapest verified option as of July 10, 2026. CompactifAI offers $0.10 in / $0.14 out (cheapest blended output price), but verify availability and SLA before committing production traffic. For workloads above ~$400–700/month in API spend, self-hosting the open weights on GPU infrastructure may be more cost-effective.

What does the Llama 4 Community License mean for my app?

The Llama 4 Community License allows free commercial use and self-hosting for most applications. The key restriction: platforms with more than 700 million monthly active users must obtain a separate license agreement directly from Meta. For the vast majority of apps and APIs, this restriction does not apply. If your platform is approaching or exceeding this threshold, contact Meta directly for the enterprise license.

Why is Llama 4 Scout faster on Groq than other hosts?

Groq uses custom LPU (Language Processing Unit) hardware designed specifically for transformer inference. As of July 10, 2026, Artificial Analysis benchmarks place Llama 4 Scout on Groq at approximately 446 tokens per second — the fastest among surveyed hosts. The trade-off is a tighter context cap (128K) and slightly higher per-token pricing ($0.11/$0.34 vs DeepInfra's $0.08/$0.30).

We build AI apps that don't hit rate limits

Retry, backoff & caching built in
Multi-provider fallback routing
Fixed price, you own the code

Get a free estimate

30-min call. No commitment.

Llama 4 Scout API Rate Limits, Pricing & Performance (July 2026)

Rate limits by tier

Token pricing

Run your own numbers

meta-llama/Llama-4-Scout-17B-16E-Instruct (host-dependent) vs the alternatives

Hitting a 429? The playbook

How to raise your limits

Free (Groq)

Pay-as-you-go (DeepInfra)

Pay-as-you-go (Fireworks)

Amazon Bedrock

Self-host

Cut your token spend

Groq for latency-critical tasks, DeepInfra for cost-critical tasks

Use DeepInfra (320K) or Fireworks (~1M) for contexts above 128K

Use Scout instead of Maverick for 80%+ of tasks

Amazon Bedrock batch for bulk non-real-time inference

Cache system prompts at the application layer

Consider CompactifAI for lowest blended cost (verify first)

Self-host for full 10M context

Frequently asked questions

We build AI apps that don't hit rate limits

Your next step

Still weighing your options?

We put the rapid in RapidDev

Llama 4 Scout API Rate Limits, Pricing & Performance (July 2026)

Rate limits by tier

Token pricing

Run your own numbers

meta-llama/Llama-4-Scout-17B-16E-Instruct (host-dependent) vs the alternatives

Hitting a 429? The playbook

How to raise your limits

Free (Groq)

Pay-as-you-go (DeepInfra)

Pay-as-you-go (Fireworks)

Amazon Bedrock

Self-host

Cut your token spend

Groq for latency-critical tasks, DeepInfra for cost-critical tasks

Use DeepInfra (320K) or Fireworks (~1M) for contexts above 128K

Use Scout instead of Maverick for 80%+ of tasks

Amazon Bedrock batch for bulk non-real-time inference

Cache system prompts at the application layer

Consider CompactifAI for lowest blended cost (verify first)

Self-host for full 10M context

Related models

Llama 4 Maverick API Rate Limits & Pricing

Llama 3.1 API Rate Limits & Pricing

Llama 3 API Rate Limits & Pricing

More meta-llama/Llama-4-Scout-17B-16E-Instruct (host-dependent) resources

Llama 4 Scout via OpenAI-Compatible Wrapper in n8n, Make & Zapier

Llama 4 Scout 429 Errors & Context Window Truncation Fixes

Frequently asked questions

We build AI apps that don't hit rate limits

Your next step

Still weighing your options?

We put the rapid in RapidDev