The 1 Million Token Context Window — What It Actually Means for You

"So should I wait for local models to get better before committing to anything?"

I get some version of this question almost every week. And honestly? Six months ago I didn't have a clean answer. Today I do.

But let's back up. Because something happened last week that changed a few of my assumptions — and it's worth unpacking carefully, because the headline is almost misleading.

Anthropic quietly made the 1 million token context window standard for Opus and Sonnet. Not a beta. Not an API-only thing. Standard.

And before you scroll past thinking "okay, bigger window, cool" — that's actually the wrong story.

The real story is that context rot might finally have a solution.

Wait, What's Context Rot?

Here's the thing nobody explains properly. Having a large context window and being able to use a large context window are two completely different things.

For the past year or so, going past 100,000–200,000 tokens in any model was basically a trap. Sure, technically you had more budget. But performance tanked. The model would lose track of things, repeat itself, miss details from earlier in the conversation. Developers call this context rot — the quality degrades as the window fills up.

So we all developed workarounds. In tools like Claude Code, the rule of thumb became: clear at around 100,000–120,000 tokens, reset the context, and start fresh. Otherwise you'd get outputs you couldn't trust.

A study by the team at Chroma Research from last summer basically confirmed what developers were experiencing in the wild. Massive drop-offs across every model as input tokens increased. It wasn't subtle. It was a cliff.

The Numbers That Actually Matter

Anthropic released benchmark data from something called the eight-needle test — a variation of the classic "needle in a haystack" problem. The idea is roughly this: you fill up the context window with a long conversation, scatter specific pieces of information throughout it (the "needles"), and then ask the model to retrieve them accurately at different points. With 1 million tokens, you're testing whether the model can actually hold and use a massive amount of context without losing track of things.

Here's where it gets interesting.

Opus 4.6 scores 78.3 on this test at 1 million tokens. For comparison: GPT-4.1 lands around 36, Gemini 3.1 Pro around 26. And Sonnet 4.5 — the previous generation — was at 18.5.

The drop from the 256k mark all the way to 1 million? About 14%. Over 750,000 additional tokens.

That's not context rot. That's genuinely usable long-context performance.

And it changes the math. Instead of clearing every 100K tokens out of necessity, you now have actual wiggle room. If you're working with a large codebase and can't afford to lose the thread, you don't have to play hacky games to keep your context artificially small. You can let it breathe.

Why I'm Not Buying a Mac Studio M3 Ultra

This brings me to the local model question. Because here's my honest read on the situation.

There's a certain appeal to the idea of running everything locally. No API costs, no dependency on someone else's infrastructure, full control. I get it. And the hardware has gotten genuinely impressive — a Mac Studio M3 Ultra can run some models that would've required a data center three years ago.

But I'm not doing it. And I'd tell most of my clients the same thing.

The gap between what you can run locally on a €15,000 machine and what Anthropic is shipping right now? It's not closing fast enough to justify the investment. These benchmark numbers — 78.3 on an eight-needle test at 1 million tokens — that's not something you're running locally anytime soon. The traction is too fast. By the time you've amortized that Mac Studio, the cloud models will have lapped themselves twice.

More importantly: Anthropic is acting like a company focused on the product, not the brand. The fact that developers at Google, Meta, and Microsoft are openly using Claude for serious coding work isn't a coincidence. It's because the model keeps actually getting better at the things developers need. That's a different dynamic than the AI hype cycle most people are riding.

What This Means for Clients: Stop Signing Long-Term AI Contracts

One thing I keep telling clients, and I'll say it here too: don't commit to 12-year subscriptions on anything meaningfully connected to AI infrastructure right now.

I don't mean don't use AI tools. Use them. Use them heavily.

I mean don't lock yourself into something based on the assumption that the current capabilities and pricing are fixed. They're not. The context window situation is a good example — the practical limits of working with AI tools changed materially in a single week. What makes sense at €X/month today may look completely different in 18 months, in either direction.

Build with current tools. Stay flexible on commitments. That's not caution for its own sake — it's just being realistic about how fast the ground is moving.

The 1 million token window isn't the story. The story is that we might finally have a model that can hold a large context and stay coherent through it. If the numbers hold up in practice — and anecdotally, they seem to — that's a meaningful change in how AI-assisted development actually works day to day.

Whether you're building something yourself or working with someone like me on web development, understanding what these tools can actually do (versus what the marketing says) matters. It shapes what's realistic to build, what's overengineered, and what advice you should probably ignore.

I can't promise you that the benchmarks translate cleanly to your specific use case. What I can offer: I keep up with this stuff obsessively so you don't have to. Let's talk if that sounds useful.