Anthropic released Claude Opus 5 on 24 July. One million tokens of context, up to 128,000 tokens of output, at 5 dollars per million input and 25 per million output.
The interesting number is the context window, and not for the reason people usually give.
What a million tokens is for
The standard demo is dropping an entire codebase into a prompt. Useful for engineering, less relevant to us.
Here is the agency version. A year of campaign data across four platforms, the client’s brand guidelines, the last twelve monthly reports, and the transcript of the quarterly review, all in one context. No retrieval layer, no chunking strategy, no vector database to maintain.
That last part is the real shift. Before, getting a model to reason across a client’s full history meant building infrastructure. Someone had to chunk documents, choose an embedding model, tune retrieval and then debug why the model kept missing the one paragraph that mattered. A large enough window makes a category of engineering work optional.
Optional, not obsolete. Retrieval still wins on cost at scale, and it still wins when the corpus is genuinely large. But for a single client account, the calculation flipped.
Worth remembering. A bigger window removes a retrieval problem. It does not remove a data quality problem. Contradictory inputs still produce confident nonsense, just across more pages.
Where the cost lands
At 5 dollars per million input tokens, filling the full window once costs about 5 dollars. That sounds trivial until you do it in a loop.
An agent that re-reads the full context on every step of a ten-step workflow spends 50 dollars on input alone, for one run, for one client. Multiply by an account list and a daily cadence, and the model that felt cheap becomes the largest line in your tooling budget.
This is the practical constraint that replaces the old one. We used to design around context limits. Now we design around context economics. The discipline is the same: put in what the task needs, not everything you have.
AI FOR YOUR TEAM
Put the new models to work on real client accounts
We train marketing teams on the workflows that hold up in production, and we build the ones that need engineering.
Three things we tested in the first week
Abstract capability claims are cheap. These are the three we actually ran, with what came out.
The full-year account review. We gave it twelve months of campaign data across four platforms plus the meeting notes, and asked what changed and why. It found a pattern we had missed: a drop in one channel that coincided with a landing page change nobody had connected to it, because the two events lived in different reports and were eight weeks apart. That is the kind of link a summarised context loses.
The brand voice check across a content batch. Twenty articles in one context, with the client’s guidelines, asked to flag where the tone drifts. This worked less well than expected. It identifies obvious violations reliably and is much weaker on the subtle drift that a good editor catches, which is the part we actually wanted help with.
The reporting draft with the raw exports. No pre-processing, just the platform exports and a request for a first draft. Good structure, correct arithmetic, and confidently wrong on one metric because two exports used different attribution windows and it picked one without saying so. That failure is worth remembering: the model does not know your data is inconsistent, it just resolves the inconsistency silently.
The pattern across all three is that a large context helps most where the value comes from spotting connections across a long period, and helps least where the task needs judgement about quality.
Does this kill retrieval
Short answer: no, and the framing is wrong.
Retrieval still wins on three counts. Cost, when the same corpus is queried repeatedly, since you pay for a fraction of it each time rather than the whole thing. Scale, when the corpus genuinely exceeds a million tokens, which is most document repositories of any age. And traceability, since a retrieval system can show you which source produced an answer, while a full-context read gives you an answer that came from everywhere at once.
What changed is the threshold. A year of one client’s marketing data no longer justifies building infrastructure. A five-year archive across forty clients still does.
Our rule now: if the whole corpus fits comfortably and gets read once or twice, put it in the context. If it gets read every day, or does not fit, keep the retrieval layer.
What we changed in our own workflow
Three things, immediately.
Quarterly reviews now get the full year in one pass. Previously we summarised each month, then summarised the summaries, and lost the detail that made a pattern visible. Now the whole period goes in at once, and the model spots the thing that happened in March and repeated in September.
We stopped maintaining a retrieval layer for single-account analysis. It was infrastructure that existed to work around a limit that no longer applies at this scale.
We tightened what goes into the context, not loosened it. This is counterintuitive and it is the important one. A larger window makes it tempting to paste everything. Everything includes the contradictory exports, the deprecated naming convention and the report nobody trusts. The model will read all of it and average across it.
A bigger context window is not permission to stop thinking about what you put in it. It is a bigger opportunity to put in the wrong thing. MyDigipal
The boundary that has not moved
Everything we said about delegation still applies. The model prepares, a person confirms anything irreversible. Publishing, sending, spending, deleting.
Opus 5 is better at long autonomous stretches, which makes that boundary more important rather than less. A model that works for twenty minutes without checking in produces more work between review points, so the review has to be deliberate rather than a glance.
We covered this when Opus 4.8 introduced dynamic workflows, and the principle held through the upgrade. What changed is the volume of output you are reviewing, not the need to review it.
If you are wiring models into workflows that touch client data or ad spend, the security controls we described in June matter more at this capability level, not less.
Sources: Claude Opus 5 versus Fable 5, benchmarks and pricing - Claude Opus 5 benchmarks, pricing and guide - Anthropic Claude model release timeline
