Charlie Hills 🦩 recently shared something that caught my attention: "BREAKING: Claude Opus 4.6 just dropped. Why you should use it instead of ChatGPT." He followed that with the real headline: Claude’s biggest weakness was its context window, and now "That’s gone now" with a 1M token context window.

That framing resonates because most "which model is best" debates ignore the daily bottleneck that actually determines output quality: how much of your real world information you can keep in the model’s working memory at once. Charlie’s post is short, punchy, and clearly written for practitioners, not spectators. I want to expand on what he’s pointing at: what a 1M token context window changes, where it still falls short, and how to use it well if your work involves long-form writing, research, or code.

"For strategic thinking. For long-form writing. For deep research. For coding. Nothing touches it right now." - Charlie Hills 🦩

The core upgrade: context is the product

Charlie described the old pain plainly: "Conversations got cut short. Documents got lost halfway through." If you have ever pasted a long brief, a stack of interview transcripts, a PRD, or a messy set of notes into a model, you know the failure mode.

A larger context window is not just a bigger input box. It changes what workflows are possible:

You can keep sources, constraints, and outputs in the same conversation without constantly re-summarizing.
You can iterate on a long draft without the model forgetting the beginning.
You can do multi-step reasoning across a wider evidence base (when you prompt for it correctly).

In practice, this reduces the "prompt tax" you pay to maintain continuity. It also reduces silent drift, where the model slowly stops respecting constraints because it no longer sees them.

What does "1M tokens" actually mean?

Charlie’s examples were concrete: "Reads an entire book in one go" and "Process your whole codebase at once." Those are good mental models, but it helps to translate tokens into decisions:

1) Fewer brittle summaries

With smaller contexts, you summarize aggressively. Summaries are lossy. They remove nuance, caveats, and edge cases. With a bigger window, you can keep raw material (notes, transcripts, specs, customer emails) alongside your working draft.

2) Better cross-referencing and consistency

Long-form writing and product work both require internal consistency: terminology, claims, dates, and positioning. Bigger context lets you ask the model to:

identify contradictions across sections
enforce a style guide across a full draft
keep a single source of truth for definitions

3) Stronger "editor" behavior

Models become more useful when they can behave like an editor with full visibility. That means you can request:

a structural critique that references exact passages
a list of unsupported claims with citations back to your notes
rewrite options that preserve the same argument but improve clarity

The practical win is not that the model "knows more." It is that it can "see more" of what you already have.

Where Claude Opus 4.6 still loses (and why that matters)

Charlie also pointed out two real constraints:

"The API is still very expensive."
"Gemini handle images and video better."

Let’s unpack both in a way that helps you choose the right tool, rather than picking a side.

Cost: context is not free

A 1M token window invites you to paste everything. That can get expensive quickly, especially via API. The trick is to treat full-context runs as premium operations, not default behavior.

A practical cost-aware approach:

Use smaller context for quick ideation and outlines.
Use full context for milestone passes (final synthesis, consistency checks, deep research summaries, large refactors).
Cache stable documents (style guide, product glossary, brand voice) in a reusable format so you do not re-send them constantly.

If you are in a team setting, set guardrails: when to run "big context," who approves it, and how to store outputs so you do not repeat work.

Multimodality: images and video are a separate game

If your workflow depends on screenshots, diagrams, whiteboards, UI walkthroughs, or video analysis, Charlie’s Gemini point is important. Text-only excellence does not automatically translate to visual understanding.

A simple division of labor often works best:

Use a model that excels at vision to extract structured notes from images/video.
Feed those notes (plus any critical frames as text descriptions) into Opus for deep synthesis, strategy, and writing.

When Opus 4.6 is the right choice

Charlie’s list is a strong starting point. Here is how I’d make it actionable.

Strategic thinking: turn context into decisions

Strategy work benefits from breadth: market notes, customer feedback, internal constraints, competitor messaging, and your current positioning.

Try this workflow:

Paste your raw inputs (bullets, notes, call transcripts, links as summaries).
Ask for a decision memo with: options, trade-offs, risks, and a recommendation.
Force specificity: "Cite the exact inputs you used for each claim."

This reduces the model’s tendency to produce generic advice and makes it auditable.

Long-form writing: fewer rewrites, stronger coherence

Large context helps most when you are beyond the outline stage. A practical approach:

Put your working draft, research notes, and voice guidelines in the same thread.
Ask for a "coherence pass" (structure, transitions, repetition).
Ask for a "truth pass" (highlight claims that need sources).
Ask for a "voice pass" (tone, sentence length, banned phrases).

With long context, you can treat the model like a real editor who has read the whole manuscript.

Deep research: synthesis over search

A bigger context window does not magically browse the web. The advantage is what happens after you collect sources: the synthesis.

If you already have PDFs, transcripts, or long notes, you can ask for:

a map of themes with supporting excerpts
a list of disagreements across sources
a set of testable hypotheses and what evidence would confirm them

The key is to separate gathering from reasoning. Gather carefully. Then let the model reason across the full set.

Coding: the promise and the limits

Charlie’s "Process your whole codebase at once" is the dream. In practice, be careful: codebases include generated files, vendored dependencies, and irrelevant noise.

A better pattern:

Provide a curated slice: the modules related to the change, plus interface definitions.
Ask for a plan first (files to touch, tests to add, risks).
Only then request patches.

Large context helps with cross-file reasoning, but you still need good engineering discipline: tests, diffs, and incremental changes.

Prompting patterns that take advantage of big context

If you want to "get the most out of Opus 4.6" (as Charlie teased), here are a few patterns that reliably improve results:

1) Provide a table of contents for your input

When you paste a lot of material, add a short index first: "Section A: brief," "Section B: notes," etc. Then tell the model how to reference sections.

2) Ask for an evidence-backed output

Add: "For each major claim, quote the supporting passage or cite the section name." This keeps outputs grounded in your inputs.

3) Use staged outputs

Request: "First output an outline, then wait." Or: "Ask me 5 questions before drafting." Big context plus staged control reduces rambling.

4) Define what not to do

Include constraints like: "Do not invent metrics," "Do not add new features," "If missing info, say what is missing." It sounds basic, but it scales with context.

So, should you switch from ChatGPT to Claude?

Charlie’s post makes a clear personal claim: he left ChatGPT for Claude because the context limitation was holding him back, and Opus 4.6 removes that blocker.

My take is slightly more tool-agnostic: switch when your bottleneck is continuity across large inputs. If your work is mostly short chats, quick brainstorming, or heavy image/video interpretation, you may not feel the upgrade as strongly.

But if your day is documents, drafts, and decisions, the ability to keep more of your world in the conversation is not a minor feature. It is the difference between "helpful" and "integrated."

P.S. Charlie ended with a good question: "Which AI model are you using the most right now?" I’d answer: whichever one best matches the job. And increasingly, that job is "reason across everything I already have."

This blog post expands on a viral LinkedIn post by Charlie Hills 🦩, I help you (actually) use AI for content.. View the original LinkedIn post →