Yesterday I wrote about Escaping Compacting Hell and how amazing the 1M context model is. I said something like “the cost is a bit steep, but the sense of freedom is hard to put a price on.”

So today I went all in. Anthropic had given me $50 in free credits.

The result? Burned through all of it in 3-4 hours.

The $50 Free Credit

I had received $50 in free credits from Anthropic. Honestly, $50 felt like a lot. My regular subscription is $20/month. Surely $50 would last me a few days, right?

Since 1M context is 5x the standard 200K, I figured the cost would be maybe 5-10x more per query. At that rate, $50 should cover a day or two of solid usage.

…That calculation was dead wrong.

1M Isn’t Just “10x More”

Let me start with the raw experience. 10x? No. It’s a different dimension entirely.

I fed it a 200MB PDF document. The whole thing. In one shot. No interruptions.

I had it analyze around 100 images, each about 5MB. All in the same session. No breaks, no hiccups.

I threw document analysis and generation tasks at it like crazy… and Compacting never triggered. Not once.

Before, loading just 3-4 PDFs would bring up “Compacting conversation…” and wipe out your context. With 1M, it felt like there was practically no limit.

This Smoothness Is Dangerous

I’ll be honest — at first, I was just having a blast.

Normally I’d be doing mental gymnastics: “If I send this, will it blow the context window?” That calculation was second nature. With 1M, none of that. Just throw tasks at it. Read this PDF. Analyze this image. Generate this document. It handled everything without flinching.

This is how AI coding tools should work, right? It always felt backwards that I had to manage the AI’s context capacity while delegating work to it.

But this smoothness… it was a little scary. Because it made me forget that the meter was running.

Something Feels Off

About two hours in, I checked my credit balance on a whim.

Wait — it’s dropped that much already?

At first, I barely noticed the cost. A few dozen cents per query. “That’s fine,” I thought. But as time went on, the drain got noticeably faster.

When you think about it, this makes perfect sense.

1M context means the conversation history doesn’t get pruned. No compression. And if nothing gets thrown away? Every single query includes the entire previous conversation as tokens. The longer your conversation, the more tokens snowball into each query.

In other words, cost doesn’t scale linearly. It scales exponentially.

The Final Moments of $50

Around the three-hour mark, things got real.

Every question felt like $2-3 vanishing from my balance. That first hour? I probably spent less than $10. The last hour? Easily over $20.

The first $10 lasted over an hour. The last $10 felt like it was gone in under 20 minutes.

That’s the terrifying thing about exponential growth. It starts gentle, almost flat. Then at some point it goes vertical. And by the time you hit that inflection point, it’s already too late.

End result: $50 gone in 3-4 hours. Clean zero.

Why This Happens

Here’s how the cost structure plays out:

Early session (0-1 hr): Small context → cents per query → smooth sailing
Mid session (1-2 hrs): Context accumulating → ~$1 per query → starting to notice
Late session (2-3 hrs): Context explosion → $2-3 per query → you can hear the meter ticking

The old 200K model used Compacting to prune older conversation history. Annoying, yes, but it kept costs relatively flat. The 1M model doesn’t prune — and you pay for that privilege.

Both approaches have trade-offs. But if you don’t understand this and just ride the “wow, so smooth!” wave… you’re in for a billing shock.

Using It Wisely

I wrote about AI subscription costs a while back. The bottom line: a tool should be used like a tool.

The 1M model is not an “always-on” mode.

When you genuinely need to process massive files. When you need to feed it a 200MB PDF in one go. When you need to analyze hundreds of images in a single session. That’s when you switch to 1M and power through.

Once the job is done? Switch back to the default model immediately.

That’s the most sensible approach. As I mentioned yesterday, the Switch Model button makes it a one-click toggle. Flip to 1M for heavy lifting, flip back when you’re done. Stick to this and you’ll save a significant amount.

One more thing: don’t stretch a single session too long. Costs grow exponentially as context accumulates. Once you’ve accomplished your goal, wrap up the session and start fresh. That alone makes a big difference.

Wrapping Up

The 1M context model is genuinely great. The freedom from Compacting hell, as I wrote yesterday — that’s real.

But if you get drunk on that freedom and leave it running all day, your wallet will be the first thing that gets liberated.

Yesterday I wrote “I hope 1M becomes the default someday.” After experiencing the costs firsthand… that day feels a bit further off than I thought. But the direction is clear. It’s expensive now, but prices will come down.

Until then, be smart about it. Turn it on when you need it, turn it off when you don’t. Follow that one simple rule and you can absolutely reap the benefits of 1M context.

If you ever get a free credit opportunity, definitely give it a shot. Once you experience that smoothness, it’s hard to go back. Just please — check your credit balance along the way.