DEV Community

Cover image for “Free Tier Eligible” Is the Most Expensive Phrase on AWS
Priyanshu Tiwari
Priyanshu Tiwari

Posted on Originally published at Medium

“Free Tier Eligible” Is the Most Expensive Phrase on AWS

t3.micro vs t3.small vs m7i-flex.large: how a ‘free’ server was burning $2.90 a day doing almost nothing, and the $19 setup I run now.

Three days after I moved my side project to AWS, I opened the billing page expecting to see $100.

It said $91.28.

Nobody had visited the site in any real numbers. Nothing had crashed. The server had mostly been sitting there, and it had still eaten almost nine dollars of my “free” credits. At that speed, the whole $100 would be gone before Christmas.

So I opened the EC2 console and looked at the instance type: m7i-flex.large. Eight gigabytes of RAM for a project with a handful of users.

I hadn’t really chosen it. When I set the server up, the AI coding agent doing the setup picked it with very simple logic: it’s the biggest machine the free plan lets you launch. Free plan, biggest box, done. I didn’t question it, because the console literally said free tier eligible.

That one word, “eligible,” was costing me about $87 a month.

What I found that evening was worse than a pricing mistake, though. The big machine was hiding a bug: my app was holding 3.9 GB of RAM while doing nothing. I’ll get to that.

And two days later, a bill from my other AWS account, the one with no free plan, showed me what big machines cost when nothing stops them: $326.50 in 17 days.

If you’re comparing t3.micro vs t3.small for a side project, or you just signed up for the AWS free plan and you’re staring at the instance list, here’s what I learned in one evening.

How the AWS free plan actually works (2026)

AWS changed the free tier in July 2025. If your account was created after July 15, 2025:

  • You get $100 in credits at sign-up, and you can earn up to $100 more by doing a few starter activities.
  • The free plan ends after 6 months or when the credits run out, whichever comes first.
  • These EC2 types are free tier eligible: t3.micro, t3.small, t4g.micro, t4g.small, c7i-flex.large and m7i-flex.large.

That last line is the trap. “Eligible” means you’re allowed to launch it on the free plan. It doesn’t mean it’s free. Every hour still burns credits at that machine’s normal price. A bigger eligible machine just burns them faster.

(Older accounts, created before that date, have the old 12-month free tier with t2.micro or t3.micro only.)

No free plan, no ceiling: +3,233%

The free plan has one thing going for it: a ceiling. When the credits or the 6 months run out, the free plan ends. AWS doesn’t start charging your card unless you upgrade to a paid plan yourself.

My other AWS account has no ceiling. It’s a normal paid account, and this month I ran a few big 16-vCPU machines on it for a compute-heavy experiment. On November 17, the billing page looked like this:

My other AWS account's billing page on November 17

  • Month-to-date: $326.50. The same 17 days last month: $9.80. That’s the +3,233% at the top.
  • Forecast for the month: $605.68, against $80.63 for all of October.
  • Almost the whole bar is one service: EC2 compute.
  • AWS’s own Cost Anomaly Detection had flagged 14 cost anomalies worth $294.61.

Nothing broke and nobody hacked me. I chose those machines on purpose. But big machines cost big money every hour they’re switched on, whether they’re busy or not, and on a paid account nothing stops the meter. AWS spotted the anomalies on its own. The bill doesn’t wait for you to look.

That’s why the rest of this article is about picking the smallest machine that works, and why I now open the billing page before I open the instance list.

The one rule: size for your heaviest moment, not your traffic

Don’t size a server for your visitors. Size it for the heaviest single thing it ever has to do. For most side projects that isn’t serving pages. It’s one of these:

  • Building the site on deploy. npm run build on a big frontend can easily want more than 1 GB of RAM.
  • One heavy background job. Anything that loads an AI model, processes images or video, or runs a headless browser.
  • A database on the same box. Postgres alone wants its share.

My site itself is mostly static pages. Serving them is light. But every deploy builds the site on the server, and one background job loads an AI model into memory. Those two moments decide the size. Not the traffic.

t3.micro vs t3.small: the actual difference

Both are burstable machines with 2 vCPUs. The differences that matter:

  • t3.micro: 1 GB RAM, baseline CPU 10% per vCPU. What it fit for me: the site, not the build or the heavy job.
  • t3.small: 2 GB RAM, baseline CPU 20% per vCPU. What it fit for me: everything, with swap.
  • m7i-flex.large: 8 GB RAM, baseline CPU 40%, full speed most of the time (Flex). What it fit for me: everything, with lots of room.

Why t3.micro didn’t work for me: with 1 GB of RAM, the build on deploy runs out of memory, and so does the heavy job. And T3 machines slow down to their baseline CPU once their CPU credits run out, so a long job on a micro would crawl.

Why t3.small did: 2 GB covers the normal load, and swap covers the short spikes (a deploy, a heavy job). The site doesn’t need more than that until real traffic shows up.

What each one costs

Prices change by region, so check yours on the EC2 pricing page. In my region (Sydney, ap-southeast-2), roughly:

  • m7i-flex.large: about $87 a month, or about $2.90 a day. $100 of credits lasts about one month.
  • t3.small: about $19 a month for the machine, around $27 with the disk and the public IP. The same credits last about 3–4 months.

Two hidden costs to remember: the EBS disk (you pay per GB per month even when the machine is stopped) and the public IPv4 address (AWS charges for those by the hour now).

The switch: two minutes of downtime, five changes

1. Change the instance type. You have to stop the machine first:

aws ec2 stop-instances --instance-ids i-xxxxxxxx
aws ec2 wait instance-stopped --instance-ids i-xxxxxxxx
aws ec2 modify-instance-attribute --instance-id i-xxxxxxxx --instance-type "{\"Value\": \"t3.small\"}"
aws ec2 start-instances --instance-ids i-xxxxxxxx
Enter fullscreen mode Exit fullscreen mode

My IP address stayed the same because it’s an Elastic IP. If yours is an auto-assigned public IP, it will change when you stop the machine, so update your DNS.

2. Set CPU credits to “standard”. T3 machines launch in unlimited mode by default, which means if the machine works hard for a long time, AWS bills you for the extra CPU. On standard mode it just slows down instead. On a budget, slow is better than a surprise bill:

aws ec2 modify-instance-credit-specification \
  --instance-credit-specification "InstanceId=i-xxxxxxxx,CpuCredits=standard"
Enter fullscreen mode Exit fullscreen mode

3. Add 4 GB of swap. This is what makes 2 GB enough. Swap is slower than RAM, but it stops a short spike from killing the build:

sudo fallocate -l 4G /swapfile
sudo chmod 600 /swapfile
sudo mkswap /swapfile
sudo swapon /swapfile
echo '/swapfile none swap sw 0 0' | sudo tee -a /etc/fstab
Enter fullscreen mode Exit fullscreen mode

Check it with free -h.

4. Use a smaller model for the heavy job. I switched the background job to a smaller AI model while traffic is low. The results are slightly less accurate, and I can switch back when it’s worth paying for.

5. Look for what you don’t use. A local Postgres is still running on the box even though the app uses a hosted database. That’s RAM for nothing, and it’s next on my list to switch off.

The 3.9 GB of RAM doing nothing

Here’s the part I promised.

Before the switch, I checked memory on the big machine. The app was holding 3.9 GB of RAM, with no job running and almost no one on the site.

It took a minute to see why. One background job had loaded the big AI model, and when the job finished, the model just… stayed. Parked in memory, forever, waiting for a job that might come tomorrow.

On an 8 GB machine I never noticed. There was so much room that a 4 GB leftover was invisible. The extra RAM wasn’t helping my app. It was hiding a problem in it.

After the switch, with the smaller model, the box sat at 493 MB in use. Same app, same site, about an eighth of the memory.

That’s the lesson I didn’t expect: a bigger server doesn’t only cost more. It hides your mistakes. A small one forces you to find them.

Deploys still build on the box, with swap catching the spike. I have $91.28 of credit left, and it now lasts months instead of weeks.

Which one should you pick?

  • If your project is a static site or a tiny API with no build step on the server, t3.micro (or t4g.micro if your code runs on ARM) is enough.
  • If you build on the server or run one heavy job (AI models, image processing, a headless browser), start with t3.small + 4 GB swap + standard CPU credits.
  • If you have real traffic or several heavy jobs at once, that’s when m7i-flex.large or bigger makes sense. Move up when the numbers say so, not before.

But honestly, don’t trust any comparison, including mine. Run your heaviest job once, watch free -h and top while it runs, and size the server for that moment. Then check memory again an hour later, when nothing is running. That second number is the one that surprised me.

Quick answers

Is t3.micro enough for a Node.js app?

For serving a small app, often yes. For running npm run build on the same machine, usually not: 1 GB runs out. Build somewhere else (your laptop or CI) or use a t3.small.

Can the AWS free plan charge my card?

Not while you stay on the free plan. When the credits or the 6 months run out, the free plan ends, and you only get billed if you upgrade to a paid plan. On a paid account there’s no such stop, so set a budget alert on day one.

Is m7i-flex.large free on the AWS free plan?

It’s free tier eligible, so you can launch it with credits. It still costs about $85–90 a month in credits, so $100 lasts roughly a month.

What does CPU credits “standard” vs “unlimited” mean?

Unlimited lets a T3 machine run above its baseline CPU and bills you for the extra. Standard caps it at the baseline once credits run out. Standard means no surprise charges.

Does adding swap to EC2 slow it down?

Only when it’s actually used. For short spikes it’s fine. If the machine swaps all the time, you need more RAM.

Can I change the instance type without losing data?

Yes. Stop it, change the type, start it. The EBS disk stays. Only an auto-assigned public IP changes.

If you’re building something on a small budget too, I post what breaks and what it costs on X: @priyanshudotsol.


Originally published on Medium.

Top comments (0)