# Replicate startup deal: $500 in credits

> Run open-source ML models in the cloud, deploy Llama, Stable Diffusion, and custom models via API without GPU management.

- **Offer:** $500 in credits
- **Save up to:** $500
- **Category:** AI & Data
- **How to get it:** Claim details are unlocked with SaaSOffers Premium. https://saasoffers.tech/offers/replicate
- **Eligibility:** AI startup using open-source models

## What you get

- $500 in Replicate credits
- Run any open-source ML model via API
- No GPU infrastructure to manage
- Auto-scaling inference
- Custom model deployment

## The SaaSOffers Replicate deal

**Offer:** $500 in credits on Replicate.

**Who qualifies:** an AI startup using open-source models.

There is no promo code to enter. This is a premium deal, so you open it from a SaaSOffers Premium account.

![Replicate homepage, $500 in credits for startups](https://zlcbkfmbwwroglopnrhf.supabase.co/storage/v1/object/public/blog-images/deals/replicate/tool1-hd.webp)

We checked these plans and prices against Replicate's official pricing and help pages in November 2026.

Replicate runs machine learning models in the cloud behind an API. You call a model such as FLUX, DeepSeek R1 or Llama 3 with one request, Replicate starts the hardware, and you get the output back. You can also package your own model with Cog, Replicate's open-source tool, and push it as a private API.

Replicate is now part of Cloudflare. It announced the deal on November 17, 2025, and Cloudflare said on December 1, 2025 that it was complete. Replicate says its API is not changing and that it carries on as a distinct brand.

Replicate has no free plan and no subscription. You pay for what you run. A few models in its "Try for free" collection run free for a limited number of runs, and after that you must add billing and buy credit. We found no public startup credit program on replicate.com, so the $500 here is not something Replicate lists for everyone.

## How Replicate pricing works

There are 2 ways Replicate charges for a run:

- **By hardware time.** Most public models bill per second for the time the model is actively processing. Setup and idle time on public models is free.
- **By output.** Replicate's 100+ official models bill per unit of output, such as per image, per 1,000 tokens or per second of video. They are always warm, so there are no cold boots, and their API schema stays stable.

Private models and deployments run on hardware reserved for you. You pay for all the time instances are online, including setup and idle time. The exception is fast-booting fine-tunes, which bill only for active processing.

### Hardware prices

| Hardware | Price per second | Price per hour | GPU memory | CPU and RAM |
|---|---|---|---|---|
| CPU (Small) | $0.000025 | $0.09 | None | 1 CPU, 2GB |
| CPU | $0.000100 | $0.36 | None | 4 CPUs, 8GB |
| Nvidia T4 | $0.000225 | $0.81 | 16GB | 4 CPUs, 16GB |
| Nvidia L40S | $0.000975 | $3.51 | 48GB | 10 CPUs, 65GB |
| 2x Nvidia L40S | $0.001950 | $7.02 | 96GB | 20 CPUs, 144GB |
| Nvidia A100 80GB | $0.001400 | $5.04 | 80GB | 10 CPUs, 144GB |
| 2x Nvidia A100 80GB | $0.002800 | $10.08 | 160GB | 20 CPUs, 288GB |
| Nvidia H100 | $0.001525 | $5.49 | 80GB | 13 CPUs, 144GB |
| 8x Nvidia H100 | $0.012200 | $43.92 | Not listed | Not listed |

Replicate has no yearly plan and no monthly plan, so there is no billed-yearly or billed-monthly price. The H200 lists at the same $5.49 per hour as the H100. Larger multi-GPU setups, including multi-GPU A100, H100 and L40S plus H200 capacity, need a committed spend contract.

### Official model prices

| Model | Price |
|---|---|
| black-forest-labs/flux-1.1-pro | $0.04 per output image |
| black-forest-labs/flux-dev | $0.025 per output image |
| black-forest-labs/flux-schnell | $3.00 per 1,000 output images |
| ideogram-ai/ideogram-v3-quality | $0.09 per output image |
| recraft-ai/recraft-v3 | $0.04 per output image |
| deepseek-ai/deepseek-r1 | $3.75 per 1 million input tokens, $0.01 per 1,000 output tokens |
| wavespeedai/wan-2.1-i2v-480p | $0.09 per second of output video |
| wavespeedai/wan-2.1-i2v-720p | $0.25 per second of output video |

Replicate's billing docs give a worked example for Llama 3.1 405B Instruct: 8 input tokens cost $0.0000760 and 43 output tokens cost $0.0004085, for $0.0004845 in total.

![Replicate plans and features before the $500 in credits deal](https://zlcbkfmbwwroglopnrhf.supabase.co/storage/v1/object/public/blog-images/deals/replicate/tool2-hd.webp)

## Billing rules to know before you spend the credits

- **Prepaid credit.** Replicate's prepaid credit page says you must buy credit upfront. Purchased credit is valid for 1 year from purchase and is not refundable. Its billing page also describes accounts billed monthly in arrears.
- **Auto reload.** Optional. The minimum threshold is $5 and the minimum reload balance is $15.
- **At zero.** When your balance hits zero, new work is blocked and running infrastructure shuts down.
- **Spend limits are gone.** As of July 1, 2025, new customers cannot set a monthly spend limit. On a deployment, you can cap the number of instances instead.
- **Failed runs.** Failed runs on public and official models are not charged. Canceled time-based runs are charged up to the cancel. Canceled runs on official models may still be charged. Failed or canceled runs on private models and deployments are billed for active time.
- **Models that call models.** You pay for the root model's time plus every model it calls.

### Rate limits

Prediction creation is limited to 600 requests per minute. All other endpoints allow 3,000 requests per minute. As your balance runs low, stricter limits apply. Replicate suggests auto reload to keep the balance above $20.

This one matters for the deal. An account that has been granted credit but has no payment method on file is limited to 1 request per second and 6 requests per minute. If your $500 arrives as granted credit, add a card before you load test or launch.

### Enterprise terms

Replicate's pricing page lists what an enterprise contract adds: a dedicated account manager, priority support, higher GPU limits, performance SLAs, help with onboarding, custom models and optimizations, and volume discounts. Committed spend is also the only way to get the larger multi-GPU setups, such as 4x or 8x A100 80GB at $20.16 and $40.32 per hour. If you expect to burn through $500 in your first month, ask about these terms before you sign a long contract elsewhere.

### What the Cloudflare deal changes

For now, nothing in your code. Replicate says existing models keep running and the API stays the same. Cloudflare says it plans to bring Replicate's catalog of more than 50,000 models to Workers AI, add fine-tuning there, and let customers run custom models built with Cog. Replicate models are also coming to Cloudflare's AI Gateway. If you already run on Cloudflare Workers, watch for that option, since it may change where you want to spend the credits.

## What the deal is worth: worked examples

### Example 1: image generation with official models

$500 buys this many images at list prices:

- flux-1.1-pro at $0.04: $500 ÷ $0.04 = **12,500 images**
- flux-dev at $0.025: $500 ÷ $0.025 = **20,000 images**
- flux-schnell at $3.00 per 1,000: $500 ÷ $0.003 = **166,666 images**

A product that generates 400 flux-dev images a day spends 400 × $0.025 = $10 a day. The credits last 50 days.

### Example 2: a private custom model on dedicated hardware

You push your own model with Cog and run it as a private model. You pay for every hour it is online.

- On an L40S at $3.51 per hour: $500 ÷ $3.51 = **142 hours**, about 6 days running 24/7
- On a T4 at $0.81 per hour: $500 ÷ $0.81 = **617 hours**, about 26 days
- On an H100 at $5.49 per hour: $500 ÷ $5.49 = **91 hours**

A full 30-day month (720 hours) on one always-on L40S costs 720 × $3.51 = **$2,527.20**. The credits cover about 20% of that. Use a deployment that scales to zero, or a fast-booting fine-tune, if traffic is light.

### Example 3: a public model billed by the second

A public model that takes 10 seconds per run on an L40S costs 10 × $0.000975 = $0.00975 per run. $500 covers about **51,282 runs**. The same 10-second run on a T4 costs $0.00225, so $500 covers about 222,222 runs, if the model fits in 16GB.

## Who gets the most from Replicate
- **Startups shipping image or video features.** Official models like flux-dev bill per image with no cold boot, so you can price your own product per output.
- **Teams testing several open-source models.** You can swap from FLUX to Ideogram to Recraft by changing the model name, with no new infrastructure.
- **Founders with a fine-tuned model.** Replicate supports fine-tuning, including a FLUX image fine-tuning guide, and fast-booting fine-tunes bill only for active time.
- **ML engineers who already use Docker.** Cog packages a model in a container with a `cog.yaml` file and a `predict.py`. You test with `cog predict`, then `cog push r8.im/<user>/<model>`. Replicate documents a GitHub Actions workflow for this.
- **Apps with spiky traffic.** Deployments scale from zero to hundreds of instances, and you can set a minimum to keep some warm.

## When Replicate is the wrong pick
- **Teams that need a hard monthly spend cap.** New accounts cannot set one. Prepaid credit and instance caps are the only brakes.
- **Workloads that run a GPU 24/7.** Always-on private models bill idle time. At $3.51 per hour, an L40S costs more per hour on Replicate than on the alternatives below.
- **Teams that only call frontier closed models.** Replicate is built around open models and hosted models from labs. If you only use one closed LLM, its own API may be simpler.
- **Companies that want invoice terms on day one.** Replicate starts most accounts on prepaid credit, and that credit is non-refundable.

## Replicate compared with alternatives

| Tool | How you pay | L40S per hour | H100 per hour | Free credit (published) |
|---|---|---|---|---|
| Replicate | Prepaid credit, per second or per output | $3.51 | $5.49 | Limited free runs on select models |
| [Modal](/offers/modal) | Starter $0 + compute, Team $250 + compute per month | $0.000542/s, about $1.95 | $0.001097/s, about $3.95 | $30/month on Starter, $100/month on Team |
| [Hugging Face](/offers/hugging-face) | Inference Endpoints by the hour | $1.80 | $4.50 | $2/month Inference Providers credits on PRO ($9/month), not usable for Endpoints |
| [Together AI](/offers/together-ai) | Per token or per image, plus dedicated GPUs | Not listed | $5.49 dedicated inference ($3.99 GPU cluster) | Not specified |

**Replicate vs Modal.** Modal charges $0.000542 per second for an L40S, about $1.95 an hour, against $3.51 on Replicate. Modal's Starter plan includes $30 of compute each month for up to 3 seats. Modal is code-first: you write Python functions. Replicate gives you a catalog of ready models you can call with no setup.

**Replicate vs Hugging Face.** Hugging Face Inference Endpoints list an L40S at $1.80 an hour and an H100 at $4.50 on AWS. A 720-hour month on one L40S is $1,296 there and $2,527.20 on Replicate. Replicate wins when you pay per output on official models and do not keep hardware running.

**Replicate vs Together AI.** Together AI lists FLUX.2 [dev] at $0.0154 per image and FLUX1.1 [pro] at $0.04 per megapixel. Its dedicated inference H100 is $5.49 an hour, the same as Replicate. Its GPU Clusters list on-demand H100s at $3.99. Compare per-image prices for your exact model before you pick.

[OpenAI](/offers/openai), [Anthropic](/offers/anthropic) and [Groq](/offers/groq) also have pages on SaaSOffers if you need closed models or fast LLM inference.

## Tips for your first weeks on Replicate
1. **Start with official models.** They are always warm and bill per output. You do not need to pass a version, only `owner/name`.
2. **Add a payment method early.** Granted credit with no card means 1 request per second. That will break any real traffic test.
3. **Set auto reload above $20.** It keeps you clear of the low-balance rate limits. The minimum threshold is $5.
4. **Cap deployment instances.** Since spend limits are gone, a maximum instance count is your main way to cap a bad day.
5. **Use webhooks for long jobs.** Replicate supports webhooks for predictions and trainings, so you do not need to poll.
6. **Use the Python or Node.js guide.** Replicate has getting-started guides for both, an HTTP API and an OpenAPI schema for other languages.

## Steps to get the Replicate deal
1. Log in to your SaaSOffers Premium account and open the Replicate deal page.
2. Confirm you qualify: an AI startup using open-source models.
3. Click **Get deal** and follow the redemption steps shown on the deal page.
4. Create or sign in to your Replicate account with the same email as your SaaSOffers account.
5. Check your Replicate billing page to confirm the $500 in credits is on the account before you start paid runs.
6. Add a payment method so the low rate limit for granted credit does not apply.

## FAQ

### Is Replicate free?

No. Replicate has no free plan or subscription. Select models in its Try for free collection run free for a limited number of runs, then you must add billing and buy prepaid credit. The SaaSOffers deal gives $500 in credits to AI startups using open-source models.

### How far does $500 of Replicate credit go?

At list prices, $500 buys 12,500 flux-1.1-pro images at $0.04 each, 20,000 flux-dev images at $0.025, or about 142 hours of a dedicated L40S GPU at $3.51 per hour.

### How does Replicate charge for a run?

Most public models bill per second of active processing, for example $0.000975 per second on an L40S. Setup and idle time on public models is free. Official models bill per output, such as per image or per token. Private models and deployments bill for all time instances are online, including idle.

### Can I set a spend limit on Replicate?

Not as a new customer. Replicate stopped letting new customers set monthly spend limits on July 1, 2025. You can cap the maximum number of instances on a deployment, and prepaid credit stops new work when the balance hits zero.

### Why are my requests being throttled on granted credit?

Accounts that have granted credit but no payment method are limited to 1 request per second and 6 requests per minute. Normal limits are 600 prediction requests per minute and 3,000 for other endpoints. Add a payment method and keep the balance above $20 with auto reload.

### Can I deploy my own model on Replicate?

Yes. Package it with Cog, Replicate's open-source tool, using a cog.yaml file and a predict.py, then run cog push to r8.im/<user>/<model>. Private models run on dedicated hardware and bill for online time, except fast-booting fine-tunes, which bill only for active time.

---
Source: SaaSOffers, https://saasoffers.tech/offers/replicate
