What Claude Academy doesn't teach you

Claude Academy launched on August 20, 2026 with 355 free courses, tutorials, and use cases. It is genuinely good, it costs nothing, and if you use Claude for work you should go take it.

This is about the shape of what is left.

None of what follows is a complaint. Every gap here exists for a structural reason — a vendor cannot write it, or has no reason to, or would have to say something against interest to do it honestly. Knowing where the gaps are is more useful than pretending they are not there.

1. When to use something else

Every one of the 22 courses assumes the answer is Claude. That is correct for a Claude course. It is not correct for you.

There are real cases where Claude is the wrong pick: a high-volume classification job where a small open model at a fraction of the cost is indistinguishable in quality; a workflow already sitting inside Microsoft 365 where Copilot's integration beats a better model reached through a connector; a task where you need image generation, which Claude does not do.

Nobody at Anthropic is going to write the page that says use the other one here. See the AI platform landscape and when not to use Claude for the honest version.

2. What breaks in month six

Courses teach the path that works. That is the correct thing for a course to do — you cannot learn a product by studying its failures.

But the deployment failures are patterned and predictable, and they arrive on a schedule:

• Month 2 — the pilot team loves it, nobody else has logged in.
• Month 4 — someone built 40 Skills and 34 of them have never fired. The Cowork skills graveyard is now a real maintenance burden.
• Month 6 — adoption plateaus around 30% weekly active, and the standard advice has stopped working.
• Month 8 — an agent that worked for five months starts producing subtly wrong output and nobody can tell you why.

None of this is in the catalog, and it is where most of the work is.

3. The budget conversation

There is no course on getting money. There is no tutorial on what to do when your security team has already said no once. There is no lesson on the renewal conversation in month eleven, when finance wants to know what the last twelve months bought.

These are the conversations that determine whether the deployment survives, and they are entirely absent — reasonably so, because a vendor writing "here is how to argue with your CFO about our invoice" is a strange artifact.

We cover them in building a business case, getting IT approval, and reporting ROI.

4. What it actually costs at your volume

Pricing pages are accurate. They are not the same as cost.

The gap between list price and your bill is made of things the courses do not model: cache hit rates that collapse when you change one line of a system prompt, agents that retry silently, cold-start context reloaded on every session, effort settings left on high by default across a team of forty.

Anthropic ships good primitives here — session budgets, spend alerts, the analytics and cost controls — and does not teach you where the money goes. See agent cost control and session economics.

5. Surviving a deprecation

Anthropic announces model retirements. Claude Opus 4.1 was retired on August 5, 2026 and requests to it now return an error rather than falling back.

That announcement is not the same as help. If you built a system on a model, tuned prompts against its specific behaviour, and calibrated an eval suite to its outputs, a retirement is a migration project with a deadline someone else set. In 2026 alone, Sonnet 4 and Opus 4 retired in June, Opus 4.1 in August, and Fable 5 and Mythos 5 were suspended worldwide for nineteen days in June by government order.

There is no course on this because a course on this would be a course on the risk of building on the vendor. See when your AI model disappears and migration.

6. The organisation the software lands in

The catalog's 148 use cases teach an individual how to do a task. Almost every real deployment problem is between people.

The warehouse team routes around the agent because they do not trust it. Legal wants a review of something no one has written down yet. A senior engineer is quietly telling juniors not to use it. A VP asked for a demo and now believes the demo is the product. Somebody has to run the change management, and it is you, without authority, on top of your actual job.

The closest the catalog comes is a tutorial on writing an AI diligence statement — useful, and about disclosure ethics rather than organisational friction.

7. Evals you can defend

Evals appear in Building with the Claude API. What is missing is the harder half: building a test set that reflects real usage rather than the twelve cases you thought of on a Tuesday, deciding what score is good enough to ship, and auditing the eval suite itself when it starts passing things it should catch.

This is the most-tested skill in Agent Manager interviews and the one most operators skip. Start at how to evaluate your agents.

8. The record of what everyone else shipped

Academy covers Anthropic. That is its remit.

But a decision made in August 2026 depends on more than what Anthropic shipped — on GPT-5.6 landing in July, Gemini 3.5 Flash opening the 3.5 family in June, xAI acquiring Cursor in May, and Meta entering the cloud market in July. Vendor training will never chronicle a competitor's roadmap.

We keep a dated timeline of every significant release from Anthropic, OpenAI, Google, Meta, and Microsoft going back to 2022, currently at 130 events.

9. The job

There are four Claude certifications and no career guidance.

Nothing tells you what a Forward Deployed Engineer is hired to do, what an AI Agent Manager is accountable for, what those roles pay, or what to build so a hiring manager believes you can do the work. A certificate is not evidence. A system you shipped is.

Aaron Levie's estimate is that 500,000 to 1 million companies will hire someone to run AI internally. Anthropic sells to those companies. It does not staff them.

The pattern

Read the nine back to back and the shape is obvious. A vendor course teaches capability. Everything above is consequence — what happens when the capability meets a budget, a skeptic, a deadline, a retirement notice, and a competitor's product.

Anthropic's own stated philosophy for Academy is mindsets over features, and they are right that mindsets outlast features. The mindset still has to survive a Tuesday in month six when the agent is wrong, the CFO has questions, and the model you built on is scheduled for retirement.

That is the part we write.

Where to start: the Claude Academy course map, then whichever role track applies to you — AI Agent Manager, Forward Deployed Engineer, or IT / admin.

Skill module: the nine-gap audit

The scenario. Your Claude deployment is six months old. The pilot team liked it. Adoption is real but not what you projected. Nothing is on fire, but something is off and you cannot name it — and next month someone is going to ask you whether this is working. You have taken the product training. That is not the missing piece.

Your task. Score your own deployment against the nine gaps above, 1 to 5, where 1 means "we have not thought about this at all" and 5 means "we have a documented answer I would defend in a meeting." Use your real deployment, not a hypothetical one. It takes about twenty minutes and you should do it alone before you do it with anyone else.

Then take the two lowest scores and write one concrete next action for each, with a date. Two, not nine — a nine-item plan is a way of doing nothing.

What good looks like:

Gap	Score	Why
1. When to use something else	2	Never evaluated. We picked Claude because I had used it.
2. What breaks in month six	3	We know Skills go stale; no process for it.
3. The budget conversation	1	Renewal is in March. I have no numbers.
4. What it costs at our volume	2	I know the invoice total. I cannot break it down.
5. Surviving a deprecation	1	Two agents are pinned to a model. No migration plan.
6. The organisation	3	Warehouse team still routes around it. Known, unaddressed.
7. Evals you can defend	2	We spot-check. No test set.
8. The industry record	4	I follow releases weekly.
9. The career	3	My title still says IT Manager.

Two lowest: #3 (budget) and #5 (deprecation).

• #3 — by Feb 14: pull three months of usage by team from the analytics dashboard, and convert the two workflows finance already cares about into hours saved. One page, ahead of the March renewal.
• #5 — by Feb 28: list every pinned model ID across our agents, and test the two highest-traffic ones against the current flagship. Write down what breaks.

The common mistake. Scoring the technical gaps and quietly skipping the organisational ones. Gaps 3, 6, and 9 are the uncomfortable rows — they are about money, politics, and your own position — and they are where deployments actually die. If your two lowest scores are both technical, you probably graded 3 and 6 generously. Go back and ask what you would actually say if the CFO asked on Monday.

Why this is the exercise. No course can run it for you, because the answers are specific to your organisation and several of them are about your organisation's dysfunction. That is the whole argument of this article, applied to you.