Claude's Context Window Is 1M Now, and I Had It Wrong

There's a sentence on my own Claude guide that isn't true anymore. It says the 1M token context window is an API number and that chat tops out at 200K. I published that on August 17. I found out it was wrong this morning, reading Anthropic's support docs to write this post.

I've corrected the page. That's not the interesting part though.

What I got wrong on my own guide

Claude Opus 5 and Claude Sonnet 5 now support a 1M token context window on all paid plans when you're chatting with Claude. Not just on the API. In the chat window, on Pro, on Max, on Team.

The rest of the lineup, verified August 27, 2026:

• Opus 4.8, Opus 4.7, Opus 4.6 and Sonnet 4.6: 500K tokens on all paid plans in chat.

• Everything else, including Haiku 4.5: 200K tokens.

• The API: 1M is the default on the models that carry it, no beta header needed, billed at standard pricing.

That's a real upgrade and it landed where most people actually work, which is the chat box and not a terminal. I'm not going to pretend otherwise.

The number is not the limit

Here's the sentence that should've been the headline, straight out of Anthropic's own platform documentation:

As token count grows, accuracy and recall degrade, a phenomenon known as context rot. This makes curating what's in context just as important as how much space is available.

That's the vendor saying it, on the same page where they tell you the window is bigger now.

Anthropic did the hard engineering here and then published the caveat themselves, which is more than most vendors manage. What catches people isn't the size.

Three gates that decide your real ceiling

What you actually get depends on three things, and none of them is the number on the marketing page:

• The model you're on: a 1M window belongs to specific models, not to your plan. Paying more doesn't widen it if you're talking to Haiku.

• The surface you opened: in Cowork, Opus 5 and Sonnet 5 carry 1M, but Anthropic states plainly that "Claude Sonnet 5 automatically compacts the conversation at 500K tokens." In Claude Code, Pro users need to enable usage credits to reach the 1M window on Opus models.

• Whether code execution is switched on: this is the one nobody checks.

That last one has a boundary sentence of its own:

Code execution must be enabled for automatic context management.

Automatic context management is the thing that summarizes your earlier messages when a conversation gets close to the ceiling, so the chat can keep going instead of stopping dead. Switch code execution off and it's gone, and you're back to hitting a wall you didn't know was there.

Where the window quietly shrinks

Everything in the request counts. The system prompt, every message including tool results, images and documents, your tool definitions, and Claude's own thinking tokens. Anthropic's support docs are blunt that tools and connectors are token-intensive, and that you can't increase the fixed window size for your plan.

Two more that surprise people:

• Caching doesn't buy you room: cached prompt prefixes still occupy the context window. Prompt caching changes what you pay for those tokens, not whether they count.

• Staying alive costs you somewhere else: longer conversations that trigger automatic context management consume more of your usage limit, so the chat keeps running and the bill for it comes out of your monthly allowance.

What I do about it on a real build

One of my published builds is a research agent that reads six investing newsletters and fifteen analyst channels on a schedule and writes a single page before the market opens. It's a large-input job by design, so it's exactly the kind of system that'll be summarizing itself to stay alive.

The rule I wrote into it has nothing to do with window size. It's an honesty rule. When something is missing, say so out loud. If a source didn't load, the page says the source didn't load, and it doesn't quietly write around the gap.

That rule's there because of the way summarization goes wrong. It doesn't announce itself. A compacted conversation looks identical to one that remembers everything, right up until you ask about something that got compressed away three hours back and get a confident answer built on a summary of a summary.

What to ask when you hire somebody

If you're paying someone to build you a Claude system that runs on long inputs, three questions get you most of the way there:

• Which model and which surface does this run on: the answer decides your real window, and "Claude" isn't an answer.

• What happens when the window fills: compaction, a fresh session, retrieval, or an error. All four are legitimate. Not knowing isn't.

• How would I know something got dropped: if there's no answer to that one, you're trusting a summary you'll never see.

Ask them of me too, if we end up talking.

Sources

Every fact above was verified on August 27, 2026 against Anthropic's own pages:

• Context windows, Claude Platform Docs covers context rot, what counts toward the window, caching, and overflow behavior.

• How large is the context window on paid Claude plans, Anthropic Help Center carries the per-model and per-surface breakdown, including the 500K Cowork compaction point.

• How do usage and length limits work, Anthropic Help Center covers the code execution requirement and what long conversations cost you in usage.

• The build I mentioned is written up in full at the Claude stock research agent, with a setup you can download and run yourself, no email required.

• The wider buyer's guide this expands on is How to Use Claude in 2026.

If you want a system built around work that runs on long inputs, book a free intro call. Fifteen to twenty minutes, me and not a sales team. Consulting is $250 for the first hour and $75 an hour after that, with the hours estimated up front, and whatever gets built runs in your Claude account and belongs to you.