The companies that build the models give the training away, and the material is good. Anthropic publishes a three-course fluency track that runs 8.25 hours across 32 lessons. OpenAI runs a free learning hub with four role-based pathways and a live session calendar, including one on measuring the return on LLM spend. If you have not sent your team to either, that is your first spend, and it is zero. What the free courses cannot do is arbitrate between the assistants you are paying for, teach on documents you are not allowed to upload to a public course, or make anyone finish. Those three gaps are what paid work is actually for, and they are much smaller than most vendors want you to think.
A declaration first. SFAI Labs runs paid sessions teaching teams to work with AI assistants on their own documents and daily work, so we are one side of this comparison and we lose revenue by writing the paragraph above. We write it anyway, because a buyer who already ran the free track is a better client. When to buy capability and when to grow it in-house is worked through in our playbook on buying versus building AI enablement, and the reason a license count never stood in for capability is argued in the case against mistaking seats for skill.
What Anthropic publishes, and how long it takes
Anthropic’s learning site moved. anthropic.com/learn now redirects to academy.claude.com, which matters later. Under the heading “Develop AI Fluency” it lists three courses, each with its lesson count and running time stated on the card:
| Course | Lessons | Quizzes | Stated time |
|---|---|---|---|
| AI Fluency: Framework and foundations | 14 | 1 | 4 hours |
| AI capabilities and limitations | 13 | 1 | 3.5 hours |
| Building effective human-agent teams (beta) | 5 | 1 | 45 minutes |
That is 8.25 hours and 32 lessons in total, free, described on the page as “Free courses, tutorials, and use cases.” You do not need a paid Claude plan to take them.
The first course is the one worth reading as a buyer rather than a learner. It teaches a four-part framework the page calls Delegation, Description, Discernment, and Diligence: deciding what to hand to an assistant, saying what you want well enough to get it, judging whether what came back is any good, and checking it before it leaves the building. Strip the branding off and that is the syllabus most paid AI fluency sessions sell. It is also the right syllabus, published free. Any provider pitching a curriculum weaker than those four headings should be asked why.
The second course covers what language models cannot do, which is where every quiet failure in a rollout starts. The third is short, marked beta, and aimed at teams putting agents into shared work.
There is also a paid proctored exam. The Claude Certified Architect Foundations page shows “Purchase | $125”, and it sits on Anthropic’s partner site rather than the public academy. It is an architect-level credential pitched at engineers. If your question is whether 60 people in operations and finance should sit an exam, the answer is no.
What OpenAI Academy actually is
OpenAI Academy is not a course list. It is a hub built around four pathways, named on the page as “Apply AI at Work”, “Build with AI”, “Lead AI Adoption”, and “Teach and learn with AI”. The first is for knowledge workers and covers prompting, reviewing output, and directing agent workflows. The third walks leaders toward an AI roadmap and a draft strategy. Both map onto what a 50 to 500 person company is trying to do.
The site also carries a live events calendar, communities for small business, nonprofits, government and news organizations, and replays of past sessions. The page’s own heading is “Build the AI skills your organization needs.”
No price appears anywhere on it. Enrollment needs a free ChatGPT account, not a paid plan and not a workspace membership. The three courses most of the search results are describing are AI Foundations, Applied AI Foundations, and Agents and Workflows, each closing in a quiz. They are not the whole catalog, which is one reason the coverage cannot agree on its size: articles ranking for this topic variously put it at three, thirteen, nineteen, and twenty courses. That spread is the tell. Most of them are recycling a launch post rather than reading the site.
Alongside the Academy sits OpenAI Certifications, announced in September 2025 with a target of certifying 10 million Americans by 2030 and a jobs platform attached. Preparation is meant to happen inside ChatGPT, Walmart was named as the anchor partner, and OpenAI told the press the launch is US-only for now. Several articles ranking for this topic describe the program as if it had already shipped at scale. Treat it as announced, not as a 2026 budget line.
The certificate question, answered once
Two different things get called the same word, and OpenAI has already settled it in its own help center, which states that Academy badges are not certifications. That one line is worth more than every blog post arguing about it.
A certificate of completion says you reached the end of some courseware. Both vendors’ free material produces at most this. A certification is a proctored exam against a defined standard, scored and dated, which is what the $125 Anthropic exam is.
For a company of 50 to 500 people, neither is the thing you are buying. Nobody in your market will ask whether your operations team holds a certificate. A finance lead will ask whether the work got faster, and no credential answers that. If you want certificates because your learning platform needs a completion field, the free courses will fill it. That is a reporting convenience, not a capability.
What a vendor’s own course will not teach you
Here is the structural point, and it is why free vendor training never closes the whole question no matter how good it gets.
A course written by a model vendor teaches that vendor’s product. It will not tell your team which assistant to reach for when three are deployed. It will not help you write an acceptable-use policy that constrains its own author’s product. And it cannot run on your contracts, your client files, or your pipeline reports, because it is a public course and your documents are not public.
That third gap is the one buyers underrate. Most teams we work with grasp prompting in the abstract fine. They stall where the abstract example becomes their own renewal letter, their own board pack, their own messy spreadsheet with four years of inconsistent column names. A free course cannot hold that material, so it works the only way it can: on generic examples. The jump from generic example to live document is where adoption dies, and it is small, specific, and teachable in a few hours with the real files in the room.
If you pay for ChatGPT, Claude, and Microsoft Copilot, which most companies at this size now do by accident rather than design, your team needs a rule of thumb for routing work between them. Neither vendor will give you that rule. The same pattern shows up with Microsoft’s catalog, where the resold instructor-led Copilot course turns out to be the free modules with a facilitator attached, and with Google’s, where the free program is strong enough to be the right first spend. Every vendor teaches its own surface well and the seams between them not at all.
The free live sessions nobody budgets for
The standard argument for paid delivery is that free means recorded video and paid means a live human. That is no longer accurate, and anyone selling against free should say so.
OpenAI Academy’s calendar lists dated, free, online sessions and turns over every few weeks. Recent entries include “ChatGPT 101 for Government Professionals”, “Performance Reviews: Using ChatGPT to Write Better Evaluations”, a Codex bootcamp on team workflows, and the one that should interest whoever approves your license bill: “Impact Math: Measuring the ROI of LLMs”, which ran on October 28, 2026. Past sessions carry replays, so a date you missed is not a closed door.
A free live session on measuring the return of license spend is a real asset, and it is where most of our inbound starts. Find that replay, or whatever replaced it on the calendar, and send someone before you book a call with us. If it answers the question, you kept the budget. If not, you arrive at the next conversation with sharper questions, which makes paid work shorter and cheaper.
What the completion numbers really say
You will find a claim circulating in training-vendor marketing that self-paced courses complete at under five percent. It is cited to open online course research and the number does not appear in that research.
The real figures are better and more useful. Katy Jordan’s 2014 study in the International Review of Research in Open and Distributed Learning found an average completion rate of about 6.5 percent across the courses then available. Her 2015 follow-up widened the sample to 221 courses and found completion ranging from 0.7 percent to 52.1 percent, with a median of 12.6 percent.
Read the spread, not the median. A seventy-fold range between the worst and best course says completion is driven by design and follow-up, not by whether money changed hands. A free course with a named owner, a calendar slot, and a manager who asks about it on Friday beats a paid course nobody blocked time for. The broader arithmetic of free against paid, including the payroll hours a free rollout quietly consumes, is set out in our honest comparison of the two.
What is left to pay for
Subtract the free material honestly and a short list remains. That is what we would charge for, and nothing else.
| What remains | Why free courseware cannot do it |
|---|---|
| Practice on your own documents | Public courses cannot host your files |
| A routing rule across the assistants you already pay for | No vendor will rank a competitor’s product |
| A before-and-after a finance lead will accept | Course completion is not a business measure |
| Someone accountable for the fact that it happened | A link in an email has no owner |
| The handful of jobs in your workflow an assistant should not touch | Vendor material is not incentivized to find them |
Market rates for a facilitated session run roughly $2,000 to $15,000, set by headcount and how many functions sit in the room. Custom automation work is a separate purchase at about $25,000 to $150,000. If a provider quotes inside the first band and the agenda looks like a walkthrough of the free curriculum, you are paying for a calendar invite.
A sequence that uses both
- This week. Pick one person to own enablement by name. Not a committee. Send the Anthropic fluency track, 8.25 hours, to everyone who touches documents for a living, with a date by which it should be done.
- Week two. Have the owner and one leader take OpenAI Academy’s leadership pathway, and put whoever handles the budget through its session on measuring LLM returns, live or on replay.
- Week three. Collect what did not transfer. Ask five people to name one task they tried with an assistant and gave up on. That list is your real curriculum, and it will be shorter and sharper than any catalog.
- Week four. Decide whether that list needs a paid session. Sometimes it is three things an owner can fix in an afternoon. Sometimes it is a confidentiality question that needs a policy before anyone touches a client file, and that is worth money.
- Day ninety. Measure one thing you measured before you started. One recurring task, timed. If it did not move, the training was not the problem.
When paying would be a mistake
Do not buy a session if nobody has run the free courses. You will pay a facilitator to deliver an intro that was free, and learn nothing about where your team is stuck.
Do not buy without an owner. Every rollout we have seen fail had budget and no name attached.
Do not buy if the real question is whether to keep the licenses. That is a usage question, and a session will paper over it for a quarter.
And do not buy a credential for business users. The $125 exam is for engineers with hands-on time behind it. Nobody in accounting needs one.
A staleness test for your shortlist
Anthropic’s learning material moved to academy.claude.com, and the certification page moved onto the partner property. If a proposal still points at the old addresses, or quotes the exam at $99 or free, whoever wrote it has not read the source this quarter. Catalogs here change every few months, and a provider who has not looked recently cannot tell you what their session adds to free material.
So ask two questions. What is in the free courses that makes part of your session redundant? And what will you do with our documents that a public course cannot? Good answers to both are short and specific. For a second opinion on where your own gap sits, we run a free AI-readiness call: tell us what your team already finished and we will say plainly whether there is anything left for us to do.
Frequently asked questions
Is OpenAI Academy free, and do we need a paid ChatGPT plan?
Free. You need a ChatGPT account to enroll, not a paid plan and not a workspace seat. No price appears on the hub.
Are Anthropic’s courses really free?
Yes. The academy page describes them as free courses and tutorials, and enrollment does not depend on a paid Claude plan.
Do these courses come with a certificate?
Courses on both sides close in a quiz and can produce a completion record. OpenAI’s own help documentation says the resulting badges are not certifications, so do not plan a credentialing program around them.
How long does the free Anthropic track take?
8.25 hours across 32 lessons, per the stated times on the three course cards. Budget two half-days if you want it done rather than started.
How much is the Claude certification exam?
The Claude Certified Architect Foundations page shows $125. Secondary articles quoting free, $49.50, or $99 are repeating older launch pricing.
Should we buy training at all if both vendors give it away?
Only for the residue: practice on your own files, a rule for choosing between the assistants you pay for, a measurement a finance lead will accept, and an owner. Everything above that line is free.
Will a vendor’s course tell us which assistant to use for what?
No, and it would be strange if it did. Routing work across ChatGPT, Claude, and Copilot is the one question each vendor’s material is structurally unable to answer.
Can we skip a paid session and just assign the free courses?
Often, yes, if one named person owns it and people have calendar time to finish. What kills a free rollout is not quality. It is that nobody was accountable.
Is OpenAI Certifications something we can use now?
Treat it as announced rather than available. It was announced in September 2025 with a 2030 target and a US-only scope, and coverage has run ahead of what shipped.
What does a paid session cost if we decide we need one?
Market range is about $2,000 to $15,000, set by headcount and how many functions attend. Custom automation work is a separate purchase at roughly $25,000 to $150,000.
What to do this week
Send the free Anthropic fluency track to everyone who writes for a living, put a name against it, and send your budget holder to OpenAI Academy’s session on measuring LLM returns. Wait three weeks, then collect the list of things people tried and gave up on. That list is the only honest brief for a paid session, and if it comes back empty you have your answer and you spent nothing.
Arthur Wandzel
