The AI the US Government Blocked for Two Weeks: Should Your Company Use It?

Facts first. None of this is rumor — it is all on the record.

On 12 June 2026, Anthropic released its top-tier models, Fable 5 and Mythos 5. Three days later the US government issued an export control order citing national security authority, and access was cut off for foreign nationals entirely — including Anthropic's own foreign staff. The trigger was a report that a research team had gotten around the model's safeguards and used it to find software vulnerabilities.

On 30 June the US Department of Commerce lifted the control, after Anthropic agreed to detect security risks proactively and to work out standards with the government. The models are available normally now, in Korea included. The full sequence is documented in Anthropic's announcement.

The summary: a government judged this model powerful enough to restrict, and today it is legally open. Which is exactly why "can we use it at work?" comes up.

The practical change with a model like Fable 5 is that the range of things you can get done by simply asking has widened.

The previous generation of AI polished sentences and summarized text. This one pulls terms out of a fifty-page contract and lays them out in a table, reads spreadsheet data and drafts the report, and produces formatted documents repeatedly at a quality you can actually use at work. A person still has to review the output, but simply deleting the time it takes to produce a first draft changes how the work feels.

The important part is that none of this requires a developer. A subscription costs a few tens of thousands of KRW a month (roughly USD 20–70) and there is essentially nothing to set up. The next step is where it gets harder.

Bringing AI into a company breaks into three stages, each with a different cost and a different payoff.

Stage 1 — Experiment with individual subscriptions (tens of thousands of KRW a month): one or two people pay for a subscription and hunt for where it works. There is no real cost exposure, but the benefit stays trapped in individual skill. The people who are good at it are the only ones who benefit.

Stage 2 — Automate one workflow (several million KRW, a few thousand USD, if outsourced): pick one recurring task and build a flow with AI inside it — classifying the day's inbound email and drafting the replies, or producing the weekly report draft automatically. This requires a build, so you either do it in-house or hire it out.

Stage 3 — Connect it to internal systems (project scale): a Q&A chatbot trained on internal documents, or automation wired into your existing database. Security review comes with it, and the cost runs from several million KRW upward depending on scope.

Most companies stop at stage 1. The ones that saw a return are the ones that reached stage 2.

From where we sit, taking these projects in, here are the prices that actually change hands:

Individual and team subscriptions: 30,000–100,000 KRW a month (approx. USD 20–70), varies by tool, paid directly

Building automation for one workflow: 1,000,000–4,000,000 KRW (approx. USD 700–2,900)

Internal-document Q&A chatbot: 1,500,000–5,000,000 KRW (approx. USD 1,100–3,600)

Package covering assessment, build, and training: around 2,500,000 KRW (approx. USD 1,800)

We publish that last configuration on our fixed price list as an in-house AI adoption starter at 2,500,000 KRW (approx. USD 1,800), covering workflow analysis, tool selection, one automation build, an internal guide, and one training session. Use it as a reference point when you compare vendors.

One caution: AI subscription and API usage fees are a monthly cost separate from the build. When you take a quote, always ask what the monthly running cost will be. A vendor that will not tell you is one to skip.

1. Where your data goes — if the work involves confidential documents or customer personal information, confirm that your inputs are excluded from training (typically an enterprise plan). This is not optional.

2. Whether there is a check for when it is wrong — AI is wrong plausibly. Put a human confirmation step into the flow before anything is sent or approved. Designs built on "the AI handles all of it" turn into incidents.

3. Whether the task still exists in six months — spending money to automate work that is about to disappear is a loss. Starting with high-frequency, fixed-format work is the standard call.